diff --git a/.pylintrc b/.pylintrc index aa29d7ade..1fb193efe 100644 --- a/.pylintrc +++ b/.pylintrc @@ -9,12 +9,13 @@ extension-pkg-whitelist= fail-on= fail-under=10 ignore=CVS -ignore-paths=^repositories/.*$, +ignore-paths=/usr/lib/.*$, + ^repositories/.*$, ^extensions/.*$, ^extensions-builtin/.*$, - /usr/lib/.*$, ^modules/dml/.*$, ^modules/models/diffusion/.*$, + ^modules/xadapters/.*$, ignore-patterns= ignored-modules= jobs=0 diff --git a/CHANGELOG.md b/CHANGELOG.md index d73710c34..9895daefa 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -69,6 +69,7 @@ - enhanced theme loader - add additional debug env variables - **Fixes**: + - add variation seed to diffusers txt2img, thanks @AI-Casanova - handle extensions that install conflicting versions of packages `onnxruntime`, `opencv2-python` - installer refresh package cache on any install diff --git a/README.md b/README.md index 9fd68aec9..a9c92b20b 100644 --- a/README.md +++ b/README.md @@ -31,8 +31,6 @@ All individual features are not listed here, instead check [ChangeLog](CHANGELOG - Enhanced *Lora*/*LoCon*/*Lyco* code supporting latest trends in training - Built-in queue management - Enterprise level logging and hardened API -- Modern localization and hints engine -- Broad compatibility with existing extensions ecosystem and new extensions manager - Built in installer with automatic updates and dependency management - Modernized UI with theme support and number of built-in themes *(dark and light)* @@ -50,7 +48,7 @@ For screenshots and informations on other available themes, see [Themes Wiki](ht **SD.Next** supports two main backends: *Diffusers* and *Original*: - **Diffusers**: Based on new [Huggingface Diffusers](https://huggingface.co/docs/diffusers/index) implementation - Supports *original* SD models as well as *all* models listed below + Supports *all* models listed below This backend is set as default for new installations See [wiki article](https://github.com/vladmandic/automatic/wiki/Diffusers) for more information - **Original**: Based on [LDM](https://github.com/Stability-AI/stablediffusion) reference implementation and significantly expanded on by [A1111](https://github.com/AUTOMATIC1111/stable-diffusion-webui) @@ -116,7 +114,7 @@ Also supported are modifiers such as: > [!IMPORTANT] > - Loading any model other than standard SD 1.x / SD 2.x requires use of backend **Diffusers** > - Loading any other models using **Original** backend is not supported -> - Loading manually download model `.safetensors` files is supported for SD 1.x / SD 2.x / SD-XL models only +> - Loading manually download model `.safetensors` files is supported for specified models only (typically SD 1.x / SD 2.x / SD-XL models only) > - For all other model types, use backend **Diffusers** and use built in Model downloader or select model from Networks -> Models -> Reference list in which case it will be auto-downloaded and loaded @@ -141,7 +139,7 @@ Also supported are modifiers such as: - If you can't run us locally, try our friends at [RunDuffusion!](https://rundiffusion.com?utm_source=github&utm_medium=referral&utm_campaign=SDNext) > [!TIP] -> - Server can run without virtual environment, +> - Server can run with or without virtual environment, Recommended to use `VENV` to avoid library version conflicts with other applications > - **nVidia/CUDA** / **AMD/ROCm** / **Intel/OneAPI** are auto-detected if present and available, For any other use case such as **DirectML**, **ONNX/Olive**, **OpenVINO** specify required parameter explicitly @@ -169,7 +167,7 @@ Below is partial list of all available parameters, run `webui --help` for the fu --listen Launch web server using public IP address, default: False --auth AUTH Set access authentication like "user:pwd,user:pwd"" --autolaunch Open the UI URL in the system's default browser upon launch - --docs Mount Gradio docs at /docs, default: False + --docs Mount API docs, default: False --no-hashing Disable hashing of checkpoints, default: False --no-metadata Disable reading of metadata from models, default: False --backend {original,diffusers} force model pipeline type @@ -208,9 +206,9 @@ SD.Next comes with several extensions pre-installed: ### **Collab** -- We'd love to have additional maintainers with full admin rights. If you're interested, ping us! -- In addition to general cross-platform code, desire is to have a lead for each of the main platforms. -This should be fully cross-platform, but we'd really love to have additional contributors and/or maintainers to join and help lead the efforts on different platforms. +- We'd love to have additional maintainers (with comes with full repo rights). If you're interested, ping us! +- In addition to general cross-platform code, desire is to have a lead for each of the main platforms +This should be fully cross-platform, but we'd really love to have additional contributors and/or maintainers to join and help lead the efforts on different platforms ### **Credits** diff --git a/modules/cmd_args.py b/modules/cmd_args.py index afe08e84c..49025613c 100644 --- a/modules/cmd_args.py +++ b/modules/cmd_args.py @@ -25,7 +25,7 @@ group.add_argument("--freeze", default=os.environ.get("SD_FREEZE", False), actio group.add_argument("--auth", type=str, default=os.environ.get("SD_AUTH", None), help='Set access authentication like "user:pwd,user:pwd""') group.add_argument("--auth-file", type=str, default=os.environ.get("SD_AUTHFILE", None), help='Set access authentication using file, default: %(default)s') group.add_argument("--autolaunch", default=os.environ.get("SD_AUTOLAUNCH", False), action='store_true', help="Open the UI URL in the system's default browser upon launch") -group.add_argument('--docs', default=os.environ.get("SD_DOCS", False), action='store_true', help = "Mount Gradio docs at /docs, default: %(default)s") +group.add_argument('--docs', default=os.environ.get("SD_DOCS", False), action='store_true', help = "Mount API docs, default: %(default)s") group.add_argument('--api-only', default=os.environ.get("SD_APIONLY", False), action='store_true', help = "Run in API only mode without starting UI") group.add_argument("--api-log", default=os.environ.get("SD_APILOG", False), action='store_true', help="Enable logging of all API requests, default: %(default)s") group.add_argument("--device-id", type=str, default=os.environ.get("SD_DEVICEID", None), help="Select the default CUDA device to use, default: %(default)s") diff --git a/modules/xadapter/adapter.py b/modules/xadapter/adapter.py new file mode 100644 index 000000000..abe9bc8fe --- /dev/null +++ b/modules/xadapter/adapter.py @@ -0,0 +1,327 @@ +import torch +import torch.nn as nn +from collections import OrderedDict +from diffusers.models.embeddings import ( + TimestepEmbedding, + Timesteps, +) + + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def get_parameter_dtype(parameter: torch.nn.Module): + try: + params = tuple(parameter.parameters()) + if len(params) > 0: + return params[0].dtype + + buffers = tuple(parameter.buffers()) + if len(buffers) > 0: + return buffers[0].dtype + + except StopIteration: + # For torch.nn.DataParallel compatibility in PyTorch 1.5 + + def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]: + tuples = [(k, v) for k, v in module.__dict__.items() if torch.is_tensor(v)] + return tuples + + gen = parameter._named_members(get_members_fn=find_tensor_attributes) + first_tuple = next(gen) + return first_tuple[1].dtype + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd(dims, self.channels, self.out_channels, 3, stride=stride, padding=padding) + else: + assert self.channels == self.out_channels + from torch.nn import MaxUnpool2d + self.op = MaxUnpool2d(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class Upsample(nn.Module): + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = nn.ConvTranspose2d(self.channels, self.out_channels, 3, stride=stride, padding=1) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x, output_size): + assert x.shape[1] == self.channels + return self.op(x, output_size) + + +class Linear(nn.Module): + def __init__(self, temb_channels, out_channels): + super(Linear, self).__init__() + self.linear = nn.Linear(temb_channels, out_channels) + + def forward(self, x): + return self.linear(x) + + + +class ResnetBlock(nn.Module): + + def __init__(self, in_c, out_c, down, up, ksize=3, sk=False, use_conv=True, enable_timestep=False, temb_channels=None, use_norm=False): + super().__init__() + self.use_norm = use_norm + self.enable_timestep = enable_timestep + ps = ksize // 2 + if in_c != out_c or sk == False: + self.in_conv = nn.Conv2d(in_c, out_c, ksize, 1, ps) + else: + self.in_conv = None + self.block1 = nn.Conv2d(out_c, out_c, 3, 1, 1) + self.act = nn.ReLU() + if use_norm: + self.norm1 = nn.GroupNorm(num_groups=32, num_channels=out_c, eps=1e-6, affine=True) + self.block2 = nn.Conv2d(out_c, out_c, ksize, 1, ps) + if sk == False: + self.skep = nn.Conv2d(in_c, out_c, ksize, 1, ps) + else: + self.skep = None + + self.down = down + self.up = up + if self.down: + self.down_opt = Downsample(in_c, use_conv=use_conv) + if self.up: + self.up_opt = Upsample(in_c, use_conv=use_conv) + if enable_timestep: + self.timestep_proj = Linear(temb_channels, out_c) + + + def forward(self, x, output_size=None, temb=None): + if self.down == True: + x = self.down_opt(x) + if self.up == True: + x = self.up_opt(x, output_size) + if self.in_conv is not None: # edit + x = self.in_conv(x) + + h = self.block1(x) + if temb is not None: + temb = self.timestep_proj(temb)[:, :, None, None] + h = h + temb + if self.use_norm: + h = self.norm1(h) + h = self.act(h) + h = self.block2(h) + if self.skep is not None: + return h + self.skep(x) + else: + return h + x + + +class Adapter_XL(nn.Module): + + def __init__(self, in_channels=[1280, 640, 320], out_channels=[1280, 1280, 640], nums_rb=3, ksize=3, sk=True, use_conv=False, use_zero_conv=True, + enable_timestep=False, use_norm=False, temb_channels=None, fusion_type='ADD'): + super(Adapter_XL, self).__init__() + self.channels = in_channels + self.nums_rb = nums_rb + self.body = [] + self.out = [] + self.use_zero_conv = use_zero_conv + self.fusion_type = fusion_type + self.gamma = [] + self.beta = [] + self.norm = [] + if fusion_type == "SPADE": + self.use_zero_conv = False + for i in range(len(self.channels)): + if self.fusion_type == 'SPADE': + # Corresponding to SPADE + self.gamma.append(nn.Conv2d(out_channels[i], out_channels[i], 1, padding=0)) + self.beta.append(nn.Conv2d(out_channels[i], out_channels[i], 1, padding=0)) + self.norm.append(nn.BatchNorm2d(out_channels[i])) + elif use_zero_conv: + self.out.append(self.make_zero_conv(out_channels[i])) + else: + self.out.append(nn.Conv2d(out_channels[i], out_channels[i], 1, padding=0)) + for j in range(nums_rb): + if i==0: + # 1280, 32, 32 -> 1280, 32, 32 + self.body.append( + ResnetBlock(in_channels[i], out_channels[i], down=False, up=False, ksize=ksize, sk=sk, use_conv=use_conv, + enable_timestep=enable_timestep, temb_channels=temb_channels, use_norm=use_norm)) + # 1280, 32, 32 -> 1280, 32, 32 + elif i==1: + # 640, 64, 64 -> 1280, 64, 64 + if j==0: + self.body.append( + ResnetBlock(in_channels[i], out_channels[i], down=False, up=False, ksize=ksize, sk=sk, + use_conv=use_conv, enable_timestep=enable_timestep, temb_channels=temb_channels, use_norm=use_norm)) + else: + self.body.append( + ResnetBlock(out_channels[i], out_channels[i], down=False, up=False, ksize=ksize,sk=sk, + use_conv=use_conv, enable_timestep=enable_timestep, temb_channels=temb_channels, use_norm=use_norm)) + else: + # 320, 64, 64 -> 640, 128, 128 + if j==0: + self.body.append( + ResnetBlock(in_channels[i], out_channels[i], down=False, up=True, ksize=ksize, sk=sk, + use_conv=True, enable_timestep=enable_timestep, temb_channels=temb_channels, use_norm=use_norm)) + # use convtranspose2d + else: + self.body.append( + ResnetBlock(out_channels[i], out_channels[i], down=False, up=False, ksize=ksize, sk=sk, + use_conv=use_conv, enable_timestep=enable_timestep, temb_channels=temb_channels, use_norm=use_norm)) + + + self.body = nn.ModuleList(self.body) + if self.use_zero_conv: + self.zero_out = nn.ModuleList(self.out) + + # if self.fusion_type == 'SPADE': + # self.norm = nn.ModuleList(self.norm) + # self.gamma = nn.ModuleList(self.gamma) + # self.beta = nn.ModuleList(self.beta) + # else: + # self.zero_out = nn.ModuleList(self.out) + + + # if enable_timestep: + # a = 320 + # + # time_embed_dim = a * 4 + # self.time_proj = Timesteps(a, True, 0) + # timestep_input_dim = a + # + # self.time_embedding = TimestepEmbedding( + # timestep_input_dim, + # time_embed_dim, + # act_fn='silu', + # post_act_fn=None, + # cond_proj_dim=None, + # ) + + + def make_zero_conv(self, channels): + + return zero_module(nn.Conv2d(channels, channels, 1, padding=0)) + + @property + def dtype(self) -> torch.dtype: + """ + `torch.dtype`: The dtype of the module (assuming that all the module parameters have the same dtype). + """ + return get_parameter_dtype(self) + + def forward(self, x, t=None): + # extract features + features = [] + b, c, _, _ = x[-1].shape + if t is not None: + if not torch.is_tensor(t): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = x[0].device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + t = torch.tensor([t], dtype=dtype, device=x[0].device) + elif len(t.shape) == 0: + t = t[None].to(x[0].device) + + t = t.expand(b) + t = self.time_proj(t) # b, 320 + t = t.to(dtype=x[0].dtype) + t = self.time_embedding(t) # b, 1280 + output_size = (b, 640, 128, 128) # last CA layer output + for i in range(len(self.channels)): + for j in range(self.nums_rb): + idx = i * self.nums_rb + j + if j == 0: + if i < 2: + out = self.body[idx](x[i], temb=t) + else: + out = self.body[idx](x[i], output_size=output_size, temb=t) + else: + out = self.body[idx](out, temb=t) + if self.fusion_type == 'SPADE': + out_gamma = self.gamma[i](out) + out_beta = self.beta[i](out) + out = [out_gamma, out_beta] + else: + out = self.zero_out[i](out) + features.append(out) + + return features + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +if __name__=='__main__': + adapter = Adapter_XL(use_zero_conv=True, + enable_timestep=True, use_norm=True, temb_channels=1280, fusion_type='SPADE').cuda() + x = [torch.randn(4, 1280, 32, 32).cuda(), torch.randn(4, 640, 64, 64).cuda(), torch.randn(4, 320, 64, 64).cuda()] + t = torch.tensor([1,2,3,4]).cuda() + result = adapter(x, t=t) + for xx in result: + print(xx[0].shape) + print(xx[1].shape) + + + + diff --git a/modules/xadapter/pipeline_sd_xl_adapter.py b/modules/xadapter/pipeline_sd_xl_adapter.py new file mode 100644 index 000000000..65e04ecab --- /dev/null +++ b/modules/xadapter/pipeline_sd_xl_adapter.py @@ -0,0 +1,1462 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import warnings +import os +import PIL +from typing import Any, Callable, Dict, List, Optional, Tuple, Union +import torch +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL +from diffusers.models.attention_processor import ( + AttnProcessor2_0, + LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, + XFormersAttnProcessor, +) +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + is_accelerate_available, + is_accelerate_version, + is_invisible_watermark_available, + logging, + replace_example_docstring, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput +from modules.xadapter.adapter import Adapter_XL +from modules.xadapter.unet_adapter import UNet2DConditionModel + + +if is_invisible_watermark_available(): + from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLAdapterPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + In addition the pipeline inherits the following loading methods: + - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`] + - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`] + + as well as the following saving methods: + - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`] + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + vae_sd1_5: AutoencoderKL, + text_encoder_sd1_5: CLIPTextModel, + tokenizer_sd1_5: CLIPTokenizer, + unet_sd1_5: UNet2DConditionModel, + scheduler_sd1_5: KarrasDiffusionSchedulers, + adapter: Adapter_XL, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + vae_sd1_5=vae_sd1_5, + text_encoder_sd1_5=text_encoder_sd1_5, + tokenizer_sd1_5=tokenizer_sd1_5, + unet_sd1_5=unet_sd1_5, + scheduler_sd1_5=scheduler_sd1_5, + adapter=adapter, + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.vae_scale_factor_sd1_5 = 2 ** (len(self.vae_sd1_5.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = self.unet.config.sample_size + self.image_processor_sd1_5 = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_sd1_5) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def enable_model_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + + device = torch.device(f"cuda:{gpu_id}") + + + self.to("cpu", silence_dtype_warnings=True) + torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + model_sequence = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + model_sequence.extend([self.unet, self.vae]) + + model_sequence.extend([self.unet_sd1_5, self.vae_sd1_5, self.text_encoder_sd1_5, self.adapter]) + + hook = None + for cpu_offloaded_model in model_sequence: + _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + + # We'll offload the last model manually. + self.final_offload_hook = hook + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + # textual inversion: procecss multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1: -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds.hidden_states[-2] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt, negative_prompt_2] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + LoRAXFormersAttnProcessor, + LoRAAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_sd1_5: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + height_sd1_5: Optional[int] = None, + width_sd1_5: Optional[int] = None, + num_inference_steps: int = 50, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + latents_sd1_5: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + prompt_embeds_sd1_5: Optional[torch.FloatTensor] = None, + negative_prompt_embeds_sd1_5: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + adapter_condition_scale: Optional[float] = 1.0, + adapter_guidance_start: Union[float, List[float]] = 0.5, + denoising_start: Optional[float] = None, + adapter_type: str = "de", # "de", "en", "en_de" + fusion_guidance_scale: Optional[float] = None, + enable_time_step: bool = False + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). + Guidance rescale factor should fix overexposure when using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + height_sd1_5 = height_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + width_sd1_5 = width_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + self.check_inputs_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, height_sd1_5, width_sd1_5, callback_steps, negative_prompt, prompt_embeds, negative_prompt_embeds + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = torch.device('cuda') + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + prompt_embeds_sd1_5 = self._encode_prompt_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds_sd1_5, + negative_prompt_embeds=negative_prompt_embeds_sd1_5, + lora_scale=text_encoder_lora_scale, + ) + # todo: implement prompt_embeds for SD1.5 + + # 4. Prepare timesteps + self.scheduler_sd1_5.set_timesteps(num_inference_steps, device=device) + timesteps_sd1_5 = self.scheduler_sd1_5.timesteps + num_inference_steps_sd1_5 = num_inference_steps + + self.scheduler.set_timesteps(num_inference_steps, device=device) + + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, adapter_guidance_start, device, denoising_start=denoising_start + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + num_channels_latents_sd1_5 = self.unet_sd1_5.config.in_channels + latents_sd1_5 = self.prepare_latents_sd1_5( + batch_size * num_images_per_prompt, + num_channels_latents_sd1_5, + height_sd1_5, + width_sd1_5, + prompt_embeds_sd1_5.dtype, + device, + generator, + latents_sd1_5, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + add_time_ids = self._get_add_time_ids( + original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 7.1 Apply denoising_end + if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + latents_sd1_5_prior = latents_sd1_5.clone() + + with self.progress_bar(total=num_inference_steps_sd1_5) as progress_bar: + for i, t in enumerate(timesteps_sd1_5): + + #################### SD1.5 forward #################### + t_sd1_5 = timesteps_sd1_5[i] + + latent_model_input = torch.cat([latents_sd1_5_prior] * 2) if do_classifier_free_guidance else latents_sd1_5_prior + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + # predict the noise residual + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + return_hidden_states=False + ) + noise_pred = unet_output.sample + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_sd1_5_prior = self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5_prior, **extra_step_kwargs, return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + # call the callback, if provided + if i == len(timesteps_sd1_5) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler_sd1_5.order == 0): + progress_bar.update() + + add_noise = True if denoising_start is None else False + latents = self.prepare_xl_latents_from_sd_1_5(latents_sd1_5_prior, latent_timestep, batch_size, + num_images_per_prompt, height, width, prompt_embeds.dtype, device, + generator=generator, add_noise=add_noise) + latents_sd1_5 = self.sd1_5_add_noise(latents_sd1_5_prior, latent_timestep, generator, device, + prompt_embeds.dtype) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + # expand the latents if we are doing classifier free guidance + + #################### SD1.5 forward #################### + t_sd1_5 = timesteps_sd1_5[i] + latent_model_input = torch.cat([latents_sd1_5] * 2) if do_classifier_free_guidance else latents + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + return_hidden_states=True, + return_encoder_feature=True + ) + noise_pred = unet_output.sample + hidden_states = unet_output.hidden_states + encoder_feature = unet_output.encoder_feature + + + # adapter forward + if adapter_type == "de": + down_bridge_residuals = None + up_block_additional_residual = self.adapter(hidden_states, t=t_sd1_5 if enable_time_step else None) + for xx in range(len(up_block_additional_residual)): + up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + elif adapter_type == "en": + up_block_additional_residual = None + down_bridge_residuals = self.adapter(encoder_feature) + for xx in range(len(down_bridge_residuals)): + down_bridge_residuals[xx] = down_bridge_residuals[xx] * adapter_condition_scale + else: + dict = self.adapter(x=hidden_states, enc_x=encoder_feature) + down_bridge_residuals = dict['encoder_features'] + up_block_additional_residual = dict['decoder_features'] + for xx in range(len(up_block_additional_residual)): + up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + for xx in range(len(down_bridge_residuals)): + down_bridge_residuals[xx] = down_bridge_residuals[xx] * adapter_condition_scale + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + + if self.scheduler_sd1_5.num_inference_steps == self.scheduler_sd1_5._step_index: # VM PATCH + self.scheduler_sd1_5._step_index = 0 + latents_sd1_5 = self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5, **extra_step_kwargs, + return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + up_block_additional_residual=up_block_additional_residual, + down_bridge_residuals=down_bridge_residuals, + return_dict=False, + fusion_guidance_scale=fusion_guidance_scale + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) + + # Overrride to properly handle the loading and unloading of the additional text encoder. + def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs): + # We could have accessed the unet config from `lora_state_dict()` too. We pass + # it here explicitly to be able to tell that it's coming from an SDXL + # pipeline. + state_dict, network_alphas = self.lora_state_dict( + pretrained_model_name_or_path_or_dict, + unet_config=self.unet.config, + **kwargs, + ) + self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet) + + text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k} + if len(text_encoder_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder, + prefix="text_encoder", + lora_scale=self.lora_scale, + ) + + text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k} + if len(text_encoder_2_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_2_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder_2, + prefix="text_encoder_2", + lora_scale=self.lora_scale, + ) + + @classmethod + def save_lora_weights( + self, + save_directory: Union[str, os.PathLike], + unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + is_main_process: bool = True, + weight_name: str = None, + save_function: Callable = None, + safe_serialization: bool = True, + ): + state_dict = {} + + def pack_weights(layers, prefix): + layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers + layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()} + return layers_state_dict + + state_dict.update(pack_weights(unet_lora_layers, "unet")) + + if text_encoder_lora_layers and text_encoder_2_lora_layers: + state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder")) + state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2")) + + self.write_lora_layers( + state_dict=state_dict, + save_directory=save_directory, + is_main_process=is_main_process, + weight_name=weight_name, + save_function=save_function, + safe_serialization=safe_serialization, + ) + + def _remove_text_encoder_monkey_patch(self): + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder) + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2) + + def _encode_prompt_sd1_5( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_sd1_5) + + text_inputs = self.tokenizer_sd1_5( + prompt, + padding="max_length", + max_length=self.tokenizer_sd1_5.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_sd1_5(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer_sd1_5.batch_decode( + untruncated_ids[:, self.tokenizer_sd1_5.model_max_length - 1: -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_sd1_5.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder_sd1_5.config, + "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder_sd1_5( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder_sd1_5 is not None: + prompt_embeds_dtype = self.text_encoder_sd1_5.dtype + elif self.unet_sd1_5 is not None: + prompt_embeds_dtype = self.unet_sd1_5.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer_sd1_5) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer_sd1_5( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder_sd1_5.config, + "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder_sd1_5( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def decode_latents_sd1_5(self, latents): + warnings.warn( + "The decode_latents method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor instead", + FutureWarning, + ) + latents = 1 / self.vae_sd1_5.config.scaling_factor * latents + image = self.vae_sd1_5.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs_sd1_5( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_xl_latents_from_sd_1_5( + self, latent, timestep, batch_size, num_images_per_prompt, height, width, dtype, device, generator=None, + add_noise=True + ): + # sd1.5 latent -> img + image = self.vae_sd1_5.decode(latent / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + image = image.resize((height, width)) + # image.save('./test_img/image_sd1_5.jpg') + # input() + + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + torch.cuda.empty_cache() + + image = self.image_processor.preprocess(image) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.vae.encode(image[i: i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.vae.encode(image).latent_dist.sample(generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + def sd1_5_add_noise(self, init_latents, timestep, generator, device, dtype): + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + image = self.vae_sd1_5.decode(init_latents / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + # image.save(f'./test_img/noisy_image_sd1_5_{int(timestep)}.jpg') + + return init_latents + + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + else: + t_start = 0 + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + if denoising_start is not None: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + timesteps = list(filter(lambda ts: ts < discrete_timestep_cutoff, timesteps)) + return torch.tensor(timesteps), len(timesteps) + + return timesteps, num_inference_steps - t_start + + def prepare_latents_sd1_5(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor_sd1_5, width // self.vae_scale_factor_sd1_5) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler_sd1_5.init_noise_sigma + return latents diff --git a/modules/xadapter/pipeline_sd_xl_adapter_controlnet.py b/modules/xadapter/pipeline_sd_xl_adapter_controlnet.py new file mode 100644 index 000000000..ddb334568 --- /dev/null +++ b/modules/xadapter/pipeline_sd_xl_adapter_controlnet.py @@ -0,0 +1,1829 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import os +from typing import Any, Callable, Dict, List, Optional, Tuple, Union +import warnings + +import torch +import PIL +import numpy as np +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer + +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin +# from diffusers.models import AutoencoderKL, UNet2DConditionModel +from diffusers.models import AutoencoderKL, ControlNetModel + +from diffusers.models.attention_processor import ( + AttnProcessor2_0, + LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, + XFormersAttnProcessor, +) +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + is_accelerate_available, + is_accelerate_version, + is_invisible_watermark_available, + logging, + replace_example_docstring, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput +from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel +from modules.xadapter.adapter import Adapter_XL +from modules.xadapter.unet_adapter import UNet2DConditionModel + +if is_invisible_watermark_available(): + from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLAdapterControlnetPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + In addition the pipeline inherits the following loading methods: + - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`] + - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`] + + as well as the following saving methods: + - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`] + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + vae_sd1_5: AutoencoderKL, + text_encoder_sd1_5: CLIPTextModel, + tokenizer_sd1_5: CLIPTokenizer, + unet_sd1_5: UNet2DConditionModel, + scheduler_sd1_5: KarrasDiffusionSchedulers, + adapter: Adapter_XL, + controlnet: ControlNetModel, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + vae_sd1_5=vae_sd1_5, + text_encoder_sd1_5=text_encoder_sd1_5, + tokenizer_sd1_5=tokenizer_sd1_5, + unet_sd1_5=unet_sd1_5, + scheduler_sd1_5=scheduler_sd1_5, + adapter=adapter, + controlnet=controlnet + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.vae_scale_factor_sd1_5 = 2 ** (len(self.vae_sd1_5.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = self.unet.config.sample_size + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor_sd1_5, do_convert_rgb=True, do_normalize=False + ) + self.image_processor_sd1_5 = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_sd1_5) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def enable_model_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + + device = torch.device(f"cuda:{gpu_id}") + + if device.type != "cpu": + self.to("cpu", silence_dtype_warnings=True) + torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + model_sequence = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + model_sequence.extend([self.unet, self.vae]) + + model_sequence.extend([self.unet_sd1_5, self.vae_sd1_5, self.text_encoder_sd1_5]) + model_sequence.extend([self.controlnet, self.adapter]) + + hook = None + for cpu_offloaded_model in model_sequence: + _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + + # We'll offload the last model manually. + self.final_offload_hook = hook + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + # textual inversion: procecss multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1 : -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds.hidden_states[-2] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt, negative_prompt_2] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + LoRAXFormersAttnProcessor, + LoRAAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_sd1_5: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + height_sd1_5: Optional[int] = None, + width_sd1_5: Optional[int] = None, + image: Union[ + torch.FloatTensor, + PIL.Image.Image, + np.ndarray, + List[torch.FloatTensor], + List[PIL.Image.Image], + List[np.ndarray], + ] = None, + num_inference_steps: int = 50, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + latents_sd1_5: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + adapter_condition_scale: Optional[float] = 1.0, + guess_mode: bool = False, + control_guidance_start: Union[float, List[float]] = 0.0, + control_guidance_end: Union[float, List[float]] = 1.0, + adapter_guidance_start: Union[float, List[float]] = 0.5, + denoising_start: Optional[float] = None, + filter_scale: Optional[float] = 0.9, + filter_range: Optional[int] = 1, + fusion_guidance_scale: Optional[float] = None, + enable_time_step: bool = False, + fusion_type: Optional[str] = 'ADD', + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). + Guidance rescale factor should fix overexposure when using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + + controlnet = self.controlnet + + skip_adapter_steps = int(adapter_guidance_start * num_inference_steps) + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [ + control_guidance_end + ] + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + height_sd1_5 = height_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + width_sd1_5 = width_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + self.check_inputs_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, + image, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = torch.device('cuda') + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + global_pool_conditions = ( + controlnet.config.global_pool_conditions + if isinstance(controlnet, ControlNetModel) + else controlnet.nets[0].config.global_pool_conditions + ) + guess_mode = guess_mode or global_pool_conditions + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # prepare controlnet image + if isinstance(controlnet, ControlNetModel): + image = self.prepare_image( + image=image, + width=width_sd1_5, + height=height_sd1_5, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height_sd1_5, width_sd1_5 = image.shape[-2:] + elif isinstance(controlnet, MultiControlNetModel): + images = [] + + for image_ in image: + image_ = self.prepare_image( + image=image_, + width=width_sd1_5, + height=height_sd1_5, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + + images.append(image_) + + image = images + height_sd1_5, width_sd1_5 = image[0].shape[-2:] + else: + assert False + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + prompt_embeds_sd1_5 = self._encode_prompt_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=text_encoder_lora_scale, + ) + # todo: implement prompt_embeds for SD1.5 + + # 4. Prepare timesteps + self.scheduler_sd1_5.set_timesteps(num_inference_steps, device=device) + timesteps_sd1_5 = self.scheduler_sd1_5.timesteps + num_inference_steps_sd1_5 = num_inference_steps + + # self.scheduler.set_timesteps(num_inference_steps-skip_adapter_steps, device=device) + self.scheduler.set_timesteps(num_inference_steps, device=device) + # timesteps = self.scheduler.timesteps + + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, adapter_guidance_start, device, denoising_start=denoising_start + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + + # 5. Prepare latent variables + # if skip_adapter_steps <= 0: + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + num_channels_latents_sd1_5 = self.unet_sd1_5.config.in_channels + latents_sd1_5 = self.prepare_latents_sd1_5( + batch_size * num_images_per_prompt, + num_channels_latents_sd1_5, + height_sd1_5, + width_sd1_5, + prompt_embeds_sd1_5.dtype, + device, + generator, + latents_sd1_5, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + add_time_ids = self._get_add_time_ids( + original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 7.1 Apply denoising_end + if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + controlnet_keep = [] + for i in range(len(timesteps_sd1_5)): + keeps = [ + 1.0 - float(i / len(timesteps_sd1_5) < s or (i + 1) / len(timesteps_sd1_5) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + latents_sd1_5_prior = latents_sd1_5.clone() + + with self.progress_bar(total=num_inference_steps_sd1_5) as progress_bar: + for i, t in enumerate(timesteps_sd1_5): + #################### SD1.5 forward #################### + t_sd1_5 = timesteps_sd1_5[i] + + latent_model_input = torch.cat([latents_sd1_5_prior] * 2) if do_classifier_free_guidance else latents_sd1_5_prior + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + # Controlnet inference + if guess_mode and do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents_sd1_5_prior + control_model_input = self.scheduler_sd1_5.scale_model_input(control_model_input, t_sd1_5) + controlnet_prompt_embeds = prompt_embeds_sd1_5.chunk(2)[1] + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds_sd1_5 + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t_sd1_5, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and do_classifier_free_guidance: + # Infered ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + return_hidden_states=False + ) + noise_pred = unet_output.sample + hidden_states = unet_output.hidden_states + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_sd1_5_prior = self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5_prior, **extra_step_kwargs, return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + # call the callback, if provided + if i == len(timesteps_sd1_5) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler_sd1_5.order == 0): + progress_bar.update() + + + add_noise = True if denoising_start is None else False + latents = self.prepare_xl_latents_from_sd_1_5(latents_sd1_5_prior, latent_timestep, batch_size, + num_images_per_prompt, height, width, prompt_embeds.dtype, device, generator=generator, add_noise=add_noise) + latents_sd1_5 = self.sd1_5_add_noise(latents_sd1_5_prior, latent_timestep, generator, device, prompt_embeds.dtype) + + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps_sd1_5) < s or (i + 1) / len(timesteps_sd1_5) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + #################### SD1.5 forward #################### + t_sd1_5 = timesteps[i] + + latent_model_input = torch.cat([latents_sd1_5] * 2) if do_classifier_free_guidance else latents_sd1_5 + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + # Controlnet inference + if guess_mode and do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents_sd1_5 + control_model_input = self.scheduler_sd1_5.scale_model_input(control_model_input, t_sd1_5) + controlnet_prompt_embeds = prompt_embeds_sd1_5.chunk(2)[1] + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds_sd1_5 + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t_sd1_5, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and do_classifier_free_guidance: + # Infered ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + return_hidden_states=True, + return_encoder_feature=True + ) + noise_pred = unet_output.sample + hidden_states = unet_output.hidden_states + + # adapter forward + down_bridge_residuals = None + up_block_additional_residual = self.adapter(hidden_states) + for xx in range(len(up_block_additional_residual)): + up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_sd1_5 = self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5, **extra_step_kwargs, return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + #################### Start of SDXL forward #################### + # if i >= skip_adapter_steps: + if True: + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + # if adapter_ablation: + # noise_pred = self.unet( + # latent_model_input, + # t, + # encoder_hidden_states=prompt_embeds, + # cross_attention_kwargs=cross_attention_kwargs, + # added_cond_kwargs=added_cond_kwargs, + # return_dict=False, + # )[0] + # else: + # # noise_pred = self.unet( + # # latent_model_input, + # # t, + # # encoder_hidden_states=prompt_embeds, + # # cross_attention_kwargs=cross_attention_kwargs, + # # added_cond_kwargs=added_cond_kwargs, + # # up_block_additional_residual=up_block_additional_residual, + # # down_bridge_residuals=down_bridge_residuals, + # # return_dict=False, + # # fusion_guidance_scale=fusion_guidance_scale, + # # fusion_type=fusion_type, + # # adapter=self.adapter if fusion_type == 'SPADE' else None + # # )[0] + # noise_pred = self.unet( + # latent_model_input, + # t, + # encoder_hidden_states=prompt_embeds, + # cross_attention_kwargs=cross_attention_kwargs, + # added_cond_kwargs=added_cond_kwargs, + # up_block_additional_residual=up_block_additional_residual, + # down_bridge_residuals=down_bridge_residuals, + # return_dict=False, + # fusion_guidance_scale=fusion_guidance_scale, + # fusion_type='ADD', + # adapter=None + # )[0] + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + up_block_additional_residual=up_block_additional_residual, + down_bridge_residuals=down_bridge_residuals, + return_dict=False, + fusion_guidance_scale=fusion_guidance_scale, + fusion_type='ADD', + adapter=None + )[0] + + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + + #################### End of SDXL forward #################### + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) + + # Overrride to properly handle the loading and unloading of the additional text encoder. + def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs): + # We could have accessed the unet config from `lora_state_dict()` too. We pass + # it here explicitly to be able to tell that it's coming from an SDXL + # pipeline. + state_dict, network_alphas = self.lora_state_dict( + pretrained_model_name_or_path_or_dict, + unet_config=self.unet.config, + **kwargs, + ) + self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet) + + text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k} + if len(text_encoder_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder, + prefix="text_encoder", + lora_scale=self.lora_scale, + ) + + text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k} + if len(text_encoder_2_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_2_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder_2, + prefix="text_encoder_2", + lora_scale=self.lora_scale, + ) + + @classmethod + def save_lora_weights( + self, + save_directory: Union[str, os.PathLike], + unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + is_main_process: bool = True, + weight_name: str = None, + save_function: Callable = None, + safe_serialization: bool = True, + ): + state_dict = {} + + def pack_weights(layers, prefix): + layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers + layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()} + return layers_state_dict + + state_dict.update(pack_weights(unet_lora_layers, "unet")) + + if text_encoder_lora_layers and text_encoder_2_lora_layers: + state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder")) + state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2")) + + self.write_lora_layers( + state_dict=state_dict, + save_directory=save_directory, + is_main_process=is_main_process, + weight_name=weight_name, + save_function=save_function, + safe_serialization=safe_serialization, + ) + + def _remove_text_encoder_monkey_patch(self): + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder) + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2) + + def _encode_prompt_sd1_5( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_sd1_5) + + text_inputs = self.tokenizer_sd1_5( + prompt, + padding="max_length", + max_length=self.tokenizer_sd1_5.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_sd1_5(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer_sd1_5.batch_decode( + untruncated_ids[:, self.tokenizer_sd1_5.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_sd1_5.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder_sd1_5.config, "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder_sd1_5( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder_sd1_5 is not None: + prompt_embeds_dtype = self.text_encoder_sd1_5.dtype + elif self.unet_sd1_5 is not None: + prompt_embeds_dtype = self.unet_sd1_5.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer_sd1_5) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer_sd1_5( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder_sd1_5.config, "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder_sd1_5( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def decode_latents_sd1_5(self, latents): + warnings.warn( + "The decode_latents method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor instead", + FutureWarning, + ) + latents = 1 / self.vae_sd1_5.config.scaling_factor * latents + image = self.vae_sd1_5.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs_sd1_5( + self, + prompt, + image, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # `prompt` needs more sophisticated handling when there are multiple + # conditionings. + if isinstance(self.controlnet, MultiControlNetModel): + if isinstance(prompt, list): + logger.warning( + f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}" + " prompts. The conditionings will be fixed across the prompts." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + def prepare_latents_sd1_5(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor_sd1_5, width // self.vae_scale_factor_sd1_5) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler_sd1_5.init_noise_sigma + return latents + + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + def prepare_xl_latents_from_sd_1_5( + self, latent, timestep, batch_size, num_images_per_prompt, height, width, dtype, device, generator=None, add_noise=True + ): + # sd1.5 latent -> img + image = self.vae_sd1_5.decode(latent / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + image = image.resize((height, width)) + + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + torch.cuda.empty_cache() + + image = self.image_processor.preprocess(image) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.vae.encode(image[i: i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.vae.encode(image).latent_dist.sample(generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + else: + t_start = 0 + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + if denoising_start is not None: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + timesteps = list(filter(lambda ts: ts < discrete_timestep_cutoff, timesteps)) + return torch.tensor(timesteps), len(timesteps) + + return timesteps, num_inference_steps - t_start + + def sd1_5_add_noise(self, init_latents, timestep, generator, device, dtype): + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + return init_latents + diff --git a/modules/xadapter/pipeline_sd_xl_adapter_controlnet_img2img.py b/modules/xadapter/pipeline_sd_xl_adapter_controlnet_img2img.py new file mode 100644 index 000000000..3f3b16ebd --- /dev/null +++ b/modules/xadapter/pipeline_sd_xl_adapter_controlnet_img2img.py @@ -0,0 +1,1928 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +import os +from typing import Any, Callable, Dict, List, Optional, Tuple, Union +import warnings +import deprecate + +import torch +import PIL +import numpy as np +import torch.nn.functional as F +from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer + +from diffusers.image_processor import VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, LoraLoaderMixin, TextualInversionLoaderMixin +# from diffusers.models import AutoencoderKL, UNet2DConditionModel +from diffusers.models import AutoencoderKL, ControlNetModel + +from diffusers.models.attention_processor import ( + AttnProcessor2_0, + LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, + XFormersAttnProcessor, +) +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + is_accelerate_available, + is_accelerate_version, + is_invisible_watermark_available, + logging, + replace_example_docstring, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput +from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel +from modules.xadapter.adapter import Adapter_XL +from modules.xadapter.unet_adapter import UNet2DConditionModel + +if is_invisible_watermark_available(): + from diffusers.pipelines.stable_diffusion_xl.watermark import StableDiffusionXLWatermarker + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionXLPipeline + + >>> pipe = StableDiffusionXLPipeline.from_pretrained( + ... "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16 + ... ) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.rescale_noise_cfg +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +class StableDiffusionXLAdapterControlnetI2IPipeline(DiffusionPipeline, FromSingleFileMixin, LoraLoaderMixin): + r""" + Pipeline for text-to-image generation using Stable Diffusion XL. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the + library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.) + + In addition the pipeline inherits the following loading methods: + - *LoRA*: [`StableDiffusionXLPipeline.load_lora_weights`] + - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`] + + as well as the following saving methods: + - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`] + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. + text_encoder ([`CLIPTextModel`]): + Frozen text-encoder. Stable Diffusion XL uses the text portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically + the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant. + text_encoder_2 ([` CLIPTextModelWithProjection`]): + Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of + [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection), + specifically the + [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k) + variant. + tokenizer (`CLIPTokenizer`): + Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + tokenizer_2 (`CLIPTokenizer`): + Second Tokenizer of class + [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer). + unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + """ + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + text_encoder_2: CLIPTextModelWithProjection, + tokenizer: CLIPTokenizer, + tokenizer_2: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + vae_sd1_5: AutoencoderKL, + text_encoder_sd1_5: CLIPTextModel, + tokenizer_sd1_5: CLIPTokenizer, + unet_sd1_5: UNet2DConditionModel, + scheduler_sd1_5: KarrasDiffusionSchedulers, + adapter: Adapter_XL, + controlnet: ControlNetModel, + force_zeros_for_empty_prompt: bool = True, + add_watermarker: Optional[bool] = None, + ): + super().__init__() + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + text_encoder_2=text_encoder_2, + tokenizer=tokenizer, + tokenizer_2=tokenizer_2, + unet=unet, + scheduler=scheduler, + vae_sd1_5=vae_sd1_5, + text_encoder_sd1_5=text_encoder_sd1_5, + tokenizer_sd1_5=tokenizer_sd1_5, + unet_sd1_5=unet_sd1_5, + scheduler_sd1_5=scheduler_sd1_5, + adapter=adapter, + controlnet=controlnet + ) + self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.vae_scale_factor_sd1_5 = 2 ** (len(self.vae_sd1_5.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.default_sample_size = self.unet.config.sample_size + self.control_image_processor = VaeImageProcessor( + vae_scale_factor=self.vae_scale_factor_sd1_5, do_convert_rgb=True, do_normalize=False + ) + self.image_processor_sd1_5 = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor_sd1_5) + + add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available() + + if add_watermarker: + self.watermark = StableDiffusionXLWatermarker() + else: + self.watermark = None + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_tiling + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_tiling + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def enable_model_cpu_offload(self, gpu_id=0): + r""" + Offloads all models to CPU using accelerate, reducing memory usage with a low impact on performance. Compared + to `enable_sequential_cpu_offload`, this method moves one whole model at a time to the GPU when its `forward` + method is called, and the model remains in GPU until the next model runs. Memory savings are lower than with + `enable_sequential_cpu_offload`, but performance is much better due to the iterative execution of the `unet`. + """ + if is_accelerate_available() and is_accelerate_version(">=", "0.17.0.dev0"): + from accelerate import cpu_offload_with_hook + else: + raise ImportError("`enable_model_cpu_offload` requires `accelerate v0.17.0` or higher.") + + device = torch.device(f"cuda:{gpu_id}") + + if device.type != "cpu": + self.to("cpu", silence_dtype_warnings=True) + torch.cuda.empty_cache() # otherwise we don't see the memory savings (but they probably exist) + + model_sequence = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + model_sequence.extend([self.unet, self.vae]) + + model_sequence.extend([self.unet_sd1_5, self.vae_sd1_5, self.text_encoder_sd1_5]) + model_sequence.extend([self.controlnet, self.adapter]) + + hook = None + for cpu_offloaded_model in model_sequence: + _, hook = cpu_offload_with_hook(cpu_offloaded_model, device, prev_module_hook=hook) + + # We'll offload the last model manually. + self.final_offload_hook = hook + + def encode_prompt( + self, + prompt: str, + prompt_2: Optional[str] = None, + device: Optional[torch.device] = None, + num_images_per_prompt: int = 1, + do_classifier_free_guidance: bool = True, + negative_prompt: Optional[str] = None, + negative_prompt_2: Optional[str] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + device = device or self._execution_device + + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + # Define tokenizers and text encoders + tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2] + text_encoders = ( + [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2] + ) + + if prompt_embeds is None: + prompt_2 = prompt_2 or prompt + # textual inversion: procecss multi-vector tokens if necessary + prompt_embeds_list = [] + prompts = [prompt, prompt_2] + for prompt, tokenizer, text_encoder in zip(prompts, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, tokenizer) + + text_inputs = tokenizer( + prompt, + padding="max_length", + max_length=tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + + text_input_ids = text_inputs.input_ids + untruncated_ids = tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = tokenizer.batch_decode(untruncated_ids[:, tokenizer.model_max_length - 1: -1]) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {tokenizer.model_max_length} tokens: {removed_text}" + ) + + prompt_embeds = text_encoder( + text_input_ids.to(device), + output_hidden_states=True, + ) + + # We are only ALWAYS interested in the pooled output of the final text encoder + pooled_prompt_embeds = prompt_embeds[0] + prompt_embeds = prompt_embeds.hidden_states[-2] + + prompt_embeds_list.append(prompt_embeds) + + prompt_embeds = torch.concat(prompt_embeds_list, dim=-1) + + # get unconditional embeddings for classifier free guidance + zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt + if do_classifier_free_guidance and negative_prompt_embeds is None and zero_out_negative_prompt: + negative_prompt_embeds = torch.zeros_like(prompt_embeds) + negative_pooled_prompt_embeds = torch.zeros_like(pooled_prompt_embeds) + elif do_classifier_free_guidance and negative_prompt_embeds is None: + negative_prompt = negative_prompt or "" + negative_prompt_2 = negative_prompt_2 or negative_prompt + + uncond_tokens: List[str] + if prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt, negative_prompt_2] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = [negative_prompt, negative_prompt_2] + + negative_prompt_embeds_list = [] + for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders): + if isinstance(self, TextualInversionLoaderMixin): + negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = tokenizer( + negative_prompt, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + negative_prompt_embeds = text_encoder( + uncond_input.input_ids.to(device), + output_hidden_states=True, + ) + # We are only ALWAYS interested in the pooled output of the final text encoder + negative_pooled_prompt_embeds = negative_prompt_embeds[0] + negative_prompt_embeds = negative_prompt_embeds.hidden_states[-2] + + negative_prompt_embeds_list.append(negative_prompt_embeds) + + negative_prompt_embeds = torch.concat(negative_prompt_embeds_list, dim=-1) + + prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype, device=device) + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + pooled_prompt_embeds = pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + if do_classifier_free_guidance: + negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.repeat(1, num_images_per_prompt).view( + bs_embed * num_images_per_prompt, -1 + ) + + return prompt_embeds, negative_prompt_embeds, pooled_prompt_embeds, negative_pooled_prompt_embeds + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt=None, + negative_prompt_2=None, + prompt_embeds=None, + negative_prompt_embeds=None, + pooled_prompt_embeds=None, + negative_pooled_prompt_embeds=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt_2 is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)): + raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + elif negative_prompt_2 is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + if prompt_embeds is not None and pooled_prompt_embeds is None: + raise ValueError( + "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`." + ) + + if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None: + raise ValueError( + "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`." + ) + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def _get_add_time_ids(self, original_size, crops_coords_top_left, target_size, dtype): + add_time_ids = list(original_size + crops_coords_top_left + target_size) + + passed_add_embed_dim = ( + self.unet.config.addition_time_embed_dim * len(add_time_ids) + self.text_encoder_2.config.projection_dim + ) + expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features + + if expected_add_embed_dim != passed_add_embed_dim: + raise ValueError( + f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`." + ) + + add_time_ids = torch.tensor([add_time_ids], dtype=dtype) + return add_time_ids + + # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae + def upcast_vae(self): + dtype = self.vae.dtype + self.vae.to(dtype=torch.float32) + use_torch_2_0_or_xformers = isinstance( + self.vae.decoder.mid_block.attentions[0].processor, + ( + AttnProcessor2_0, + XFormersAttnProcessor, + LoRAXFormersAttnProcessor, + LoRAAttnProcessor2_0, + ), + ) + # if xformers or torch_2_0 is used attention block does not need + # to be in float32 which can save lots of memory + if use_torch_2_0_or_xformers: + self.vae.post_quant_conv.to(dtype) + self.vae.decoder.conv_in.to(dtype) + self.vae.decoder.mid_block.to(dtype) + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + prompt_2: Optional[Union[str, List[str]]] = None, + prompt_sd1_5: Optional[Union[str, List[str]]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + height_sd1_5: Optional[int] = None, + width_sd1_5: Optional[int] = None, + image: Union[ + torch.FloatTensor, + PIL.Image.Image, + np.ndarray, + List[torch.FloatTensor], + List[PIL.Image.Image], + List[np.ndarray], + ] = None, + source_img: Union[ + torch.FloatTensor, + PIL.Image.Image, + np.ndarray, + List[torch.FloatTensor], + List[PIL.Image.Image], + List[np.ndarray], + ] = None, + num_inference_steps: int = 50, + denoising_end: Optional[float] = None, + guidance_scale: float = 5.0, + negative_prompt: Optional[Union[str, List[str]]] = None, + negative_prompt_2: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + latents_sd1_5: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None, + callback_steps: int = 1, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + original_size: Optional[Tuple[int, int]] = None, + crops_coords_top_left: Tuple[int, int] = (0, 0), + target_size: Optional[Tuple[int, int]] = None, + controlnet_conditioning_scale: Union[float, List[float]] = 1.0, + adapter_condition_scale: Optional[float] = 1.0, + guess_mode: bool = False, + control_guidance_start: Union[float, List[float]] = 0.0, + control_guidance_end: Union[float, List[float]] = 1.0, + adapter_guidance_start: Union[float, List[float]] = 0.5, + denoising_start: Optional[float] = None, + adapter_type: str = "de", # "de", "en", "en_de" + fusion_guidance_scale: Optional[float] = None, + enable_time_step: bool = False, + fusion_type: Optional[str] = 'ADD', + ): + r""" + Function invoked when calling the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`. + instead. + prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is + used in both text-encoders + height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + denoising_end (`float`, *optional*): + When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be + completed before it is intentionally prematurely terminated. As a result, the returned sample will + still retain a substantial amount of noise as determined by the discrete timesteps selected by the + scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a + "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image + Output**](https://huggingface.co/docs/diffusers/api/pipelines/stable_diffusion/stable_diffusion_xl#refining-the-image-output) + guidance_scale (`float`, *optional*, defaults to 5.0): + Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598). + `guidance_scale` is defined as `w` of equation 2. of [Imagen + Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale > + 1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`, + usually at the expense of lower image quality. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + negative_prompt_2 (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and + `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to + [`schedulers.DDIMScheduler`], will be ignored for others. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html) + to make generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor will ge generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. + If not provided, pooled text embeddings will be generated from `prompt` input argument. + negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt` + input argument. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generate image. Choose between + [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead + of a plain tuple. + callback (`Callable`, *optional*): + A function that will be called every `callback_steps` steps during inference. The function will be + called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`. + callback_steps (`int`, *optional*, defaults to 1): + The frequency at which the `callback` function will be called. If not specified, the callback will be + called at every step. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.7): + Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of + [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). + Guidance rescale factor should fix overexposure when using zero terminal SNR. + original_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + If `original_size` is not the same as `target_size` the image will appear to be down- or upsampled. + `original_size` defaults to `(width, height)` if not specified. Part of SDXL's micro-conditioning as + explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)): + `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position + `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting + `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of + [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)): + For most cases, `target_size` should be set to the desired height and width of the generated image. If + not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in + section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952). + + Examples: + + Returns: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] or `tuple`: + [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] if `return_dict` is True, otherwise a + `tuple`. When returning a tuple, the first element is a list with the generated images. + """ + # 0. Default height and width to unet + + controlnet = self.controlnet + + skip_adapter_steps = int(adapter_guidance_start * num_inference_steps) + + # align format for control guidance + if not isinstance(control_guidance_start, list) and isinstance(control_guidance_end, list): + control_guidance_start = len(control_guidance_end) * [control_guidance_start] + elif not isinstance(control_guidance_end, list) and isinstance(control_guidance_start, list): + control_guidance_end = len(control_guidance_start) * [control_guidance_end] + elif not isinstance(control_guidance_start, list) and not isinstance(control_guidance_end, list): + mult = len(controlnet.nets) if isinstance(controlnet, MultiControlNetModel) else 1 + control_guidance_start, control_guidance_end = mult * [control_guidance_start], mult * [ + control_guidance_end + ] + + height = height or self.default_sample_size * self.vae_scale_factor + width = width or self.default_sample_size * self.vae_scale_factor + + height_sd1_5 = height_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + width_sd1_5 = width_sd1_5 or self.default_sample_size_sd1_5 * self.vae_scale_factor_sd1_5 + + original_size = original_size or (height, width) + target_size = target_size or (height, width) + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + prompt_2, + height, + width, + callback_steps, + negative_prompt, + negative_prompt_2, + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) + + self.check_inputs_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, + image, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + controlnet_conditioning_scale, + control_guidance_start, + control_guidance_end, + ) + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = torch.device('cuda') + + if isinstance(controlnet, MultiControlNetModel) and isinstance(controlnet_conditioning_scale, float): + controlnet_conditioning_scale = [controlnet_conditioning_scale] * len(controlnet.nets) + + global_pool_conditions = ( + controlnet.config.global_pool_conditions + if isinstance(controlnet, ControlNetModel) + else controlnet.nets[0].config.global_pool_conditions + ) + guess_mode = guess_mode or global_pool_conditions + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + do_classifier_free_guidance = guidance_scale > 1.0 + + # prepare controlnet image + if isinstance(controlnet, ControlNetModel): + image = self.prepare_image( + image=image, + width=width_sd1_5, + height=height_sd1_5, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + height_sd1_5, width_sd1_5 = image.shape[-2:] + elif isinstance(controlnet, MultiControlNetModel): + images = [] + + for image_ in image: + image_ = self.prepare_image( + image=image_, + width=width_sd1_5, + height=height_sd1_5, + batch_size=batch_size * num_images_per_prompt, + num_images_per_prompt=num_images_per_prompt, + device=device, + dtype=controlnet.dtype, + do_classifier_free_guidance=do_classifier_free_guidance, + guess_mode=guess_mode, + ) + + images.append(image_) + + image = images + height_sd1_5, width_sd1_5 = image[0].shape[-2:] + else: + assert False + + # 3. Encode input prompt + text_encoder_lora_scale = ( + cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None + ) + ( + prompt_embeds, + negative_prompt_embeds, + pooled_prompt_embeds, + negative_pooled_prompt_embeds, + ) = self.encode_prompt( + prompt=prompt, + prompt_2=prompt_2, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + negative_prompt_2=negative_prompt_2, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + pooled_prompt_embeds=pooled_prompt_embeds, + negative_pooled_prompt_embeds=negative_pooled_prompt_embeds, + lora_scale=text_encoder_lora_scale, + ) + + prompt_embeds_sd1_5 = self._encode_prompt_sd1_5( + prompt if prompt_sd1_5 is None else prompt_sd1_5, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt, + prompt_embeds=None, + negative_prompt_embeds=None, + lora_scale=text_encoder_lora_scale, + ) + # todo: implement prompt_embeds for SD1.5 + + # 4. Prepare timesteps + self.scheduler_sd1_5.set_timesteps(num_inference_steps, device=device) + timesteps_sd1_5 = self.scheduler_sd1_5.timesteps + num_inference_steps_sd1_5 = num_inference_steps + latent_timestep_sd1_5 = timesteps_sd1_5[:1].repeat(batch_size * num_images_per_prompt) + + self.scheduler.set_timesteps(num_inference_steps, device=device) + + timesteps, num_inference_steps = self.get_timesteps( + num_inference_steps, adapter_guidance_start, device, denoising_start=denoising_start + ) + latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt) + + # 4.1 prepare image + source_img = self.image_processor_sd1_5.preprocess(source_img).to(dtype=torch.float32) + + # 5. Prepare latent variables + # if skip_adapter_steps <= 0: + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + num_channels_latents_sd1_5 = self.unet_sd1_5.config.in_channels + + latents_sd1_5 = self.prepare_latents_sd1_5( + source_img, + latent_timestep_sd1_5, + batch_size, + num_images_per_prompt, + prompt_embeds_sd1_5.dtype, + device, + generator, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 7. Prepare added time ids & embeddings + add_text_embeds = pooled_prompt_embeds + add_time_ids = self._get_add_time_ids( + original_size, crops_coords_top_left, target_size, dtype=prompt_embeds.dtype + ) + + if do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0) + add_text_embeds = torch.cat([negative_pooled_prompt_embeds, add_text_embeds], dim=0) + add_time_ids = torch.cat([add_time_ids, add_time_ids], dim=0) + + prompt_embeds = prompt_embeds.to(device) + add_text_embeds = add_text_embeds.to(device) + add_time_ids = add_time_ids.to(device).repeat(batch_size * num_images_per_prompt, 1) + + # 8. Denoising loop + num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0) + + # 7.1 Apply denoising_end + if denoising_end is not None and type(denoising_end) == float and denoising_end > 0 and denoising_end < 1: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_end * self.scheduler.config.num_train_timesteps) + ) + ) + num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps))) + timesteps = timesteps[:num_inference_steps] + + controlnet_keep = [] + for i in range(len(timesteps_sd1_5)): + keeps = [ + 1.0 - float(i / len(timesteps_sd1_5) < s or (i + 1) / len(timesteps_sd1_5) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + latents_sd1_5_prior = latents_sd1_5.clone() + + with self.progress_bar(total=num_inference_steps_sd1_5) as progress_bar: + for i, t in enumerate(timesteps_sd1_5): + # expand the latents if we are doing classifier free guidance + + #################### SD1.5 forward #################### + # t_sd1_5 = timesteps[i] + t_sd1_5 = timesteps_sd1_5[i] + + latent_model_input = torch.cat( + [latents_sd1_5_prior] * 2) if do_classifier_free_guidance else latents_sd1_5_prior + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + # Controlnet inference + if guess_mode and do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents_sd1_5_prior + control_model_input = self.scheduler_sd1_5.scale_model_input(control_model_input, t_sd1_5) + controlnet_prompt_embeds = prompt_embeds_sd1_5.chunk(2)[1] + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds_sd1_5 + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t_sd1_5, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and do_classifier_free_guidance: + # Infered ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + return_hidden_states=False + ) + noise_pred = unet_output.sample + hidden_states = unet_output.hidden_states + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_sd1_5_prior = \ + self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5_prior, **extra_step_kwargs, + return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + # call the callback, if provided + if i == len(timesteps_sd1_5) - 1 or ( + (i + 1) > num_warmup_steps and (i + 1) % self.scheduler_sd1_5.order == 0): + progress_bar.update() + + + add_noise = True if denoising_start is None else False + latents = self.prepare_xl_latents_from_sd_1_5(latents_sd1_5_prior, latent_timestep, batch_size, + num_images_per_prompt, height, width, prompt_embeds.dtype, device, + generator=generator, add_noise=add_noise) + latents_sd1_5 = self.sd1_5_add_noise(latents_sd1_5_prior, latent_timestep, generator, device, + prompt_embeds.dtype) + + + controlnet_keep = [] + for i in range(len(timesteps)): + keeps = [ + 1.0 - float(i / len(timesteps_sd1_5) < s or (i + 1) / len(timesteps_sd1_5) > e) + for s, e in zip(control_guidance_start, control_guidance_end) + ] + controlnet_keep.append(keeps[0] if isinstance(controlnet, ControlNetModel) else keeps) + + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + + #################### SD1.5 forward #################### + t_sd1_5 = timesteps[i] + + latent_model_input = torch.cat([latents_sd1_5] * 2) if do_classifier_free_guidance else latents_sd1_5 + latent_model_input = self.scheduler_sd1_5.scale_model_input(latent_model_input, t_sd1_5) + + # Controlnet inference + if guess_mode and do_classifier_free_guidance: + # Infer ControlNet only for the conditional batch. + control_model_input = latents_sd1_5 + control_model_input = self.scheduler_sd1_5.scale_model_input(control_model_input, t_sd1_5) + controlnet_prompt_embeds = prompt_embeds_sd1_5.chunk(2)[1] + else: + control_model_input = latent_model_input + controlnet_prompt_embeds = prompt_embeds_sd1_5 + + if isinstance(controlnet_keep[i], list): + cond_scale = [c * s for c, s in zip(controlnet_conditioning_scale, controlnet_keep[i])] + else: + controlnet_cond_scale = controlnet_conditioning_scale + if isinstance(controlnet_cond_scale, list): + controlnet_cond_scale = controlnet_cond_scale[0] + cond_scale = controlnet_cond_scale * controlnet_keep[i] + + down_block_res_samples, mid_block_res_sample = self.controlnet( + control_model_input, + t_sd1_5, + encoder_hidden_states=controlnet_prompt_embeds, + controlnet_cond=image, + conditioning_scale=cond_scale, + guess_mode=guess_mode, + return_dict=False, + ) + + if guess_mode and do_classifier_free_guidance: + # Infered ControlNet only for the conditional batch. + # To apply the output of ControlNet to both the unconditional and conditional batches, + # add 0 to the unconditional batch to keep it unchanged. + down_block_res_samples = [torch.cat([torch.zeros_like(d), d]) for d in down_block_res_samples] + mid_block_res_sample = torch.cat([torch.zeros_like(mid_block_res_sample), mid_block_res_sample]) + + # predict the noise residual + unet_output = self.unet_sd1_5( + latent_model_input, + t_sd1_5, + encoder_hidden_states=prompt_embeds_sd1_5, + cross_attention_kwargs=cross_attention_kwargs, + down_block_additional_residuals=down_block_res_samples, + mid_block_additional_residual=mid_block_res_sample, + return_hidden_states=True, + return_encoder_feature=True + ) + noise_pred = unet_output.sample + hidden_states = unet_output.hidden_states + encoder_feature = unet_output.encoder_feature + + # adapter forward + # if i >= skip_adapter_steps: + if adapter_type == "de": + down_bridge_residuals = None + # up_block_additional_residual = self.adapter(hidden_states, t=t_sd1_5 if enable_time_step else None) + # if fusion_type is not "SPADE": + # for xx in range(len(up_block_additional_residual)): + # up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + up_block_additional_residual = self.adapter(hidden_states) + for xx in range(len(up_block_additional_residual)): + up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + + elif adapter_type == "en": + up_block_additional_residual = None + down_bridge_residuals = self.adapter(encoder_feature) + for xx in range(len(down_bridge_residuals)): + down_bridge_residuals[xx] = down_bridge_residuals[xx] * adapter_condition_scale + else: + dict = self.adapter(x=hidden_states, enc_x=encoder_feature) + down_bridge_residuals = dict['encoder_features'] + up_block_additional_residual = dict['decoder_features'] + for xx in range(len(up_block_additional_residual)): + up_block_additional_residual[xx] = up_block_additional_residual[xx] * adapter_condition_scale + for xx in range(len(down_bridge_residuals)): + down_bridge_residuals[xx] = down_bridge_residuals[xx] * adapter_condition_scale + + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents_sd1_5 = \ + self.scheduler_sd1_5.step(noise_pred, t_sd1_5, latents_sd1_5, **extra_step_kwargs, return_dict=False)[0] + + #################### End of SD1.5 forward #################### + + #################### Start of SDXL forward #################### + # if i >= skip_adapter_steps: + if True: + latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents + + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids} + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + cross_attention_kwargs=cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + up_block_additional_residual=up_block_additional_residual, + down_bridge_residuals=down_bridge_residuals, + return_dict=False, + fusion_guidance_scale=fusion_guidance_scale, + fusion_type='ADD', + adapter=None + )[0] + + # perform guidance + if do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond) + + if do_classifier_free_guidance and guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + #################### End of SDXL forward #################### + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + callback(i, t, latents) + + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.dtype == torch.float16 and self.vae.config.force_upcast: + self.upcast_vae() + latents = latents.to(next(iter(self.vae.post_quant_conv.parameters())).dtype) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0] + else: + image = latents + return StableDiffusionXLPipelineOutput(images=image) + + # apply watermark if available + if self.watermark is not None: + image = self.watermark.apply_watermark(image) + + image = self.image_processor.postprocess(image, output_type=output_type) + + # Offload last model to CPU + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.final_offload_hook.offload() + + if not return_dict: + return (image,) + + return StableDiffusionXLPipelineOutput(images=image) + + # Overrride to properly handle the loading and unloading of the additional text encoder. + def load_lora_weights(self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], **kwargs): + # We could have accessed the unet config from `lora_state_dict()` too. We pass + # it here explicitly to be able to tell that it's coming from an SDXL + # pipeline. + state_dict, network_alphas = self.lora_state_dict( + pretrained_model_name_or_path_or_dict, + unet_config=self.unet.config, + **kwargs, + ) + self.load_lora_into_unet(state_dict, network_alphas=network_alphas, unet=self.unet) + + text_encoder_state_dict = {k: v for k, v in state_dict.items() if "text_encoder." in k} + if len(text_encoder_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder, + prefix="text_encoder", + lora_scale=self.lora_scale, + ) + + text_encoder_2_state_dict = {k: v for k, v in state_dict.items() if "text_encoder_2." in k} + if len(text_encoder_2_state_dict) > 0: + self.load_lora_into_text_encoder( + text_encoder_2_state_dict, + network_alphas=network_alphas, + text_encoder=self.text_encoder_2, + prefix="text_encoder_2", + lora_scale=self.lora_scale, + ) + + @classmethod + def save_lora_weights( + self, + save_directory: Union[str, os.PathLike], + unet_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + text_encoder_2_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None, + is_main_process: bool = True, + weight_name: str = None, + save_function: Callable = None, + safe_serialization: bool = True, + ): + state_dict = {} + + def pack_weights(layers, prefix): + layers_weights = layers.state_dict() if isinstance(layers, torch.nn.Module) else layers + layers_state_dict = {f"{prefix}.{module_name}": param for module_name, param in layers_weights.items()} + return layers_state_dict + + state_dict.update(pack_weights(unet_lora_layers, "unet")) + + if text_encoder_lora_layers and text_encoder_2_lora_layers: + state_dict.update(pack_weights(text_encoder_lora_layers, "text_encoder")) + state_dict.update(pack_weights(text_encoder_2_lora_layers, "text_encoder_2")) + + self.write_lora_layers( + state_dict=state_dict, + save_directory=save_directory, + is_main_process=is_main_process, + weight_name=weight_name, + save_function=save_function, + safe_serialization=safe_serialization, + ) + + def _remove_text_encoder_monkey_patch(self): + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder) + self._remove_text_encoder_monkey_patch_classmethod(self.text_encoder_2) + + def _encode_prompt_sd1_5( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer_sd1_5) + + text_inputs = self.tokenizer_sd1_5( + prompt, + padding="max_length", + max_length=self.tokenizer_sd1_5.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer_sd1_5(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer_sd1_5.batch_decode( + untruncated_ids[:, self.tokenizer_sd1_5.model_max_length - 1: -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer_sd1_5.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder_sd1_5.config, + "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + prompt_embeds = self.text_encoder_sd1_5( + text_input_ids.to(device), + attention_mask=attention_mask, + ) + prompt_embeds = prompt_embeds[0] + + if self.text_encoder_sd1_5 is not None: + prompt_embeds_dtype = self.text_encoder_sd1_5.dtype + elif self.unet_sd1_5 is not None: + prompt_embeds_dtype = self.unet_sd1_5.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer_sd1_5) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer_sd1_5( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder_sd1_5.config, + "use_attention_mask") and self.text_encoder_sd1_5.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder_sd1_5( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + return prompt_embeds + + def decode_latents_sd1_5(self, latents): + warnings.warn( + "The decode_latents method is deprecated and will be removed in a future version. Please" + " use VaeImageProcessor instead", + FutureWarning, + ) + latents = 1 / self.vae_sd1_5.config.scaling_factor * latents + image = self.vae_sd1_5.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def check_inputs_sd1_5( + self, + prompt, + image, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + controlnet_conditioning_scale=1.0, + control_guidance_start=0.0, + control_guidance_end=1.0, + ): + if (callback_steps is None) or ( + callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0) + ): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + # `prompt` needs more sophisticated handling when there are multiple + # conditionings. + if isinstance(self.controlnet, MultiControlNetModel): + if isinstance(prompt, list): + logger.warning( + f"You have {len(self.controlnet.nets)} ControlNets and you have passed {len(prompt)}" + " prompts. The conditionings will be fixed across the prompts." + ) + + # Check `image` + is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance( + self.controlnet, torch._dynamo.eval_frame.OptimizedModule + ) + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + self.check_image(image, prompt, prompt_embeds) + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if not isinstance(image, list): + raise TypeError("For multiple controlnets: `image` must be type `list`") + + # When `image` is a nested list: + # (e.g. [[canny_image_1, pose_image_1], [canny_image_2, pose_image_2]]) + elif any(isinstance(i, list) for i in image): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif len(image) != len(self.controlnet.nets): + raise ValueError( + f"For multiple controlnets: `image` must have the same length as the number of controlnets, but got {len(image)} images and {len(self.controlnet.nets)} ControlNets." + ) + + for image_ in image: + self.check_image(image_, prompt, prompt_embeds) + else: + assert False + + # Check `controlnet_conditioning_scale` + if ( + isinstance(self.controlnet, ControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, ControlNetModel) + ): + if not isinstance(controlnet_conditioning_scale, float): + raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.") + elif ( + isinstance(self.controlnet, MultiControlNetModel) + or is_compiled + and isinstance(self.controlnet._orig_mod, MultiControlNetModel) + ): + if isinstance(controlnet_conditioning_scale, list): + if any(isinstance(i, list) for i in controlnet_conditioning_scale): + raise ValueError("A single batch of multiple conditionings are supported at the moment.") + elif isinstance(controlnet_conditioning_scale, list) and len(controlnet_conditioning_scale) != len( + self.controlnet.nets + ): + raise ValueError( + "For multiple controlnets: When `controlnet_conditioning_scale` is specified as `list`, it must have" + " the same length as the number of controlnets" + ) + else: + assert False + + if not isinstance(control_guidance_start, (tuple, list)): + control_guidance_start = [control_guidance_start] + + if not isinstance(control_guidance_end, (tuple, list)): + control_guidance_end = [control_guidance_end] + + if len(control_guidance_start) != len(control_guidance_end): + raise ValueError( + f"`control_guidance_start` has {len(control_guidance_start)} elements, but `control_guidance_end` has {len(control_guidance_end)} elements. Make sure to provide the same number of elements to each list." + ) + + if isinstance(self.controlnet, MultiControlNetModel): + if len(control_guidance_start) != len(self.controlnet.nets): + raise ValueError( + f"`control_guidance_start`: {control_guidance_start} has {len(control_guidance_start)} elements but there are {len(self.controlnet.nets)} controlnets available. Make sure to provide {len(self.controlnet.nets)}." + ) + + for start, end in zip(control_guidance_start, control_guidance_end): + if start >= end: + raise ValueError( + f"control guidance start: {start} cannot be larger or equal to control guidance end: {end}." + ) + if start < 0.0: + raise ValueError(f"control guidance start: {start} can't be smaller than 0.") + if end > 1.0: + raise ValueError(f"control guidance end: {end} can't be larger than 1.0.") + + # def prepare_latents_sd1_5(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + # shape = (batch_size, num_channels_latents, height // self.vae_scale_factor_sd1_5, width // self.vae_scale_factor_sd1_5) + # if isinstance(generator, list) and len(generator) != batch_size: + # raise ValueError( + # f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + # f" size of {batch_size}. Make sure the batch size matches the length of the generators." + # ) + # + # if latents is None: + # latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # else: + # latents = latents.to(device) + # + # # scale the initial noise by the standard deviation required by the scheduler + # latents = latents * self.scheduler_sd1_5.init_noise_sigma + # return latents + def prepare_latents_sd1_5(self, image, timestep, batch_size, num_images_per_prompt, dtype, device, generator=None): + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.vae_sd1_5.encode(image[i : i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.vae_sd1_5.encode(image).latent_dist.sample(generator) + + init_latents = self.vae_sd1_5.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + deprecation_message = ( + f"You have passed {batch_size} text prompts (`prompt`), but only {init_latents.shape[0]} initial" + " images (`image`). Initial images are now duplicating to match the number of text prompts. Note" + " that this behavior is deprecated and will be removed in a version 1.0.0. Please make sure to update" + " your script to pass as many initial images as text prompts to suppress this warning." + ) + deprecate("len(prompt) != len(image)", "1.0.0", deprecation_message, standard_warn=False) + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + shape = init_latents.shape + + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + + # get latents + init_latents = self.scheduler_sd1_5.add_noise(init_latents, noise, timestep) + latents = init_latents + + return latents + + def prepare_image( + self, + image, + width, + height, + batch_size, + num_images_per_prompt, + device, + dtype, + do_classifier_free_guidance=False, + guess_mode=False, + ): + image = self.control_image_processor.preprocess(image, height=height, width=width).to(dtype=torch.float32) + image_batch_size = image.shape[0] + + if image_batch_size == 1: + repeat_by = batch_size + else: + # image batch size is the same as prompt batch size + repeat_by = num_images_per_prompt + + image = image.repeat_interleave(repeat_by, dim=0) + + image = image.to(device=device, dtype=dtype) + + if do_classifier_free_guidance and not guess_mode: + image = torch.cat([image] * 2) + + return image + + def check_image(self, image, prompt, prompt_embeds): + image_is_pil = isinstance(image, PIL.Image.Image) + image_is_tensor = isinstance(image, torch.Tensor) + image_is_np = isinstance(image, np.ndarray) + image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image) + image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor) + image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray) + + if ( + not image_is_pil + and not image_is_tensor + and not image_is_np + and not image_is_pil_list + and not image_is_tensor_list + and not image_is_np_list + ): + raise TypeError( + f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}" + ) + + if image_is_pil: + image_batch_size = 1 + else: + image_batch_size = len(image) + + if prompt is not None and isinstance(prompt, str): + prompt_batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + prompt_batch_size = len(prompt) + elif prompt_embeds is not None: + prompt_batch_size = prompt_embeds.shape[0] + + if image_batch_size != 1 and image_batch_size != prompt_batch_size: + raise ValueError( + f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}" + ) + + def prepare_latents_from_noisy_latent(self, latent, device, dtype, generator, height, width, adapter_guidance_start, timesteps): + # sd1.5 noisy latent -> image + image = self.vae_sd1_5.decode(latent / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + # image = self.image_processor_sd1_5.postprocess(image, do_denormalize=do_denormalize)[0] + # image.save(f'./test_img/noisy_latent_{adapter_guidance_start:.2f}.png') + # image -> SDXL latent + image = image.resize((height, width)) + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + image = self.image_processor.preprocess(image) + image = image.to(device=device, dtype=dtype) + init_latents = self.vae.encode(image).latent_dist.sample() + init_latents = init_latents.to(dtype) + init_latents = self.vae.config.scaling_factor * init_latents + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + print(timesteps) + print(timesteps[-1]) + input() + # init_latents = self.scheduler.add_noise(init_latents, noise, timesteps[-1]) + # if self.vae.config.force_upcast: + # self.vae.to(dtype) + # image = self.vae.decode(init_latents / self.vae.config.scaling_factor, return_dict=False)[0] + # image = self.image_processor.postprocess(image, output_type="pil")[0] + # image.save(f'./test_img/noisy_init_latent_sdxl_{adapter_guidance_start:.2f}.png') + + return init_latents + + def prepare_xl_latents_from_sd_1_5( + self, latent, timestep, batch_size, num_images_per_prompt, height, width, dtype, device, generator=None, add_noise=True + ): + # sd1.5 latent -> img + image = self.vae_sd1_5.decode(latent / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + image = image.resize((height, width)) + # image.save('./test_img/image_sd1_5.jpg') + # input() + + if not isinstance(image, (torch.Tensor, PIL.Image.Image, list)): + raise ValueError( + f"`image` has to be of type `torch.Tensor`, `PIL.Image.Image` or list but is {type(image)}" + ) + + # Offload text encoder if `enable_model_cpu_offload` was enabled + if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None: + self.text_encoder_2.to("cpu") + torch.cuda.empty_cache() + + image = self.image_processor.preprocess(image) + + image = image.to(device=device, dtype=dtype) + + batch_size = batch_size * num_images_per_prompt + + if image.shape[1] == 4: + init_latents = image + + else: + # make sure the VAE is in float32 mode, as it overflows in float16 + if self.vae.config.force_upcast: + image = image.float() + self.vae.to(dtype=torch.float32) + + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + elif isinstance(generator, list): + init_latents = [ + self.vae.encode(image[i: i + 1]).latent_dist.sample(generator[i]) for i in range(batch_size) + ] + init_latents = torch.cat(init_latents, dim=0) + else: + init_latents = self.vae.encode(image).latent_dist.sample(generator) + + if self.vae.config.force_upcast: + self.vae.to(dtype) + + init_latents = init_latents.to(dtype) + init_latents = self.vae.config.scaling_factor * init_latents + + if batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] == 0: + # expand init_latents for batch_size + additional_image_per_prompt = batch_size // init_latents.shape[0] + init_latents = torch.cat([init_latents] * additional_image_per_prompt, dim=0) + elif batch_size > init_latents.shape[0] and batch_size % init_latents.shape[0] != 0: + raise ValueError( + f"Cannot duplicate `image` of batch size {init_latents.shape[0]} to {batch_size} text prompts." + ) + else: + init_latents = torch.cat([init_latents], dim=0) + + if add_noise: + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + latents = init_latents + + return latents + + def get_timesteps(self, num_inference_steps, strength, device, denoising_start=None): + # get the original timestep using init_timestep + if denoising_start is None: + init_timestep = min(int(num_inference_steps * strength), num_inference_steps) + t_start = max(num_inference_steps - init_timestep, 0) + else: + t_start = 0 + + timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :] + + # Strength is irrelevant if we directly request a timestep to start at; + # that is, strength is determined by the denoising_start instead. + if denoising_start is not None: + discrete_timestep_cutoff = int( + round( + self.scheduler.config.num_train_timesteps + - (denoising_start * self.scheduler.config.num_train_timesteps) + ) + ) + timesteps = list(filter(lambda ts: ts < discrete_timestep_cutoff, timesteps)) + return torch.tensor(timesteps), len(timesteps) + + return timesteps, num_inference_steps - t_start + + def sd1_5_add_noise(self, init_latents, timestep, generator, device, dtype): + shape = init_latents.shape + noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + # get latents + init_latents = self.scheduler.add_noise(init_latents, noise, timestep) + + image = self.vae_sd1_5.decode(init_latents / self.vae_sd1_5.config.scaling_factor, return_dict=False)[0] + do_denormalize = [True] * image.shape[0] + image = self.image_processor_sd1_5.postprocess(image, output_type='pil', do_denormalize=do_denormalize)[0] + # image.save(f'./test_img/noisy_image_sd1_5_{int(timestep)}.jpg') + + return init_latents + diff --git a/modules/xadapter/unet_adapter.py b/modules/xadapter/unet_adapter.py new file mode 100644 index 000000000..b99639466 --- /dev/null +++ b/modules/xadapter/unet_adapter.py @@ -0,0 +1,1077 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import UNet2DConditionLoadersMixin +from diffusers.utils import BaseOutput, logging +from diffusers.models.activations import get_activation +from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor +from diffusers.models.embeddings import ( + GaussianFourierProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + PositionNet, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from diffusers.models.modeling_utils import ModelMixin +from diffusers.models.unet_2d_blocks import ( + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + hidden_states: Optional[list] = None + encoder_feature: Optional[list] = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be either `UNetMidBlock2DCrossAttn` or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int` or `Tuple[int]`, *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. + conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer. + projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = list(reversed(transformer_layers_per_block)) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type == "gated": + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + self.position_net = PositionNet(positive_len=positive_len, out_dim=cross_attention_dim) + + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "set_processor"): + processors[f"{name}.processor"] = module.processor + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + self.set_attn_processor(AttnProcessor()) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + up_block_additional_residual: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + return_hidden_states: bool = False, + return_encoder_feature: bool = False, + return_early: bool = False, + down_bridge_residuals: Optional[Tuple[torch.Tensor]] = None, + fusion_guidance_scale: Optional[torch.FloatTensor] = None, + fusion_type: Optional[str] = 'ADD', + adapter: Optional = None + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + + Returns: + [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + ############## bridge usage ################## + if return_hidden_states: + hidden_states = [] + return_dict = True + ############## end of bridge usage ################## + + + + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]): + logger.info("Forward upsample size to force interpolation output size.") + forward_upsample_size = True + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) # 2, 320 + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + + if return_encoder_feature: + encoder_feature = [] + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None + is_bridge_encoder = down_bridge_residuals is not None + is_bridge = up_block_additional_residual is not None + + down_block_res_samples = (sample,) + + + + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_block_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + + if is_bridge_encoder and len(down_bridge_residuals) > 0: + sample += down_bridge_residuals.pop(0) + + if return_encoder_feature: + encoder_feature.append(sample) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb) + + if is_adapter and len(down_block_additional_residuals) > 0: + sample += down_block_additional_residuals.pop(0) + + if is_bridge_encoder and len(down_bridge_residuals) > 0: + sample += down_bridge_residuals.pop(0) + + down_block_res_samples += res_samples + + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + if return_encoder_feature and return_early: + return encoder_feature + + # 4. mid + if self.mid_block is not None: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + ################# bridge usage ################# + + if is_bridge: + if up_block_additional_residual[0].shape != sample.shape: + print('HERE SKIP1') + elif fusion_guidance_scale is not None: + sample = sample + fusion_guidance_scale * (up_block_additional_residual.pop(0) - sample) + else: + sample += up_block_additional_residual.pop(0) + ################# end of bridge usage ################# + # 5. up + + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, temb=emb, res_hidden_states_tuple=res_samples, upsample_size=upsample_size + ) + + + ################# bridge usage ################# + if is_bridge and len(up_block_additional_residual) > 0: + if sample.shape != up_block_additional_residual[0].shape: + print('HERE SKIP2') + elif fusion_guidance_scale is not None: + sample = sample + fusion_guidance_scale * (up_block_additional_residual.pop(0) - sample) + else: + sample += up_block_additional_residual.pop(0) + + if return_hidden_states and i > 0: + # Collect last three up blk in SD1.5 + hidden_states.append(sample) + ################# end of bridge usage ################# + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample, hidden_states=hidden_states if return_hidden_states else None, + encoder_feature=encoder_feature if return_encoder_feature else None) diff --git a/modules/xadapter/utils.py b/modules/xadapter/utils.py new file mode 100644 index 000000000..02990f645 --- /dev/null +++ b/modules/xadapter/utils.py @@ -0,0 +1,14 @@ +import os +import imageio +import numpy as np +from typing import Union + +import torch +import torchvision +import torch.distributed as dist + +from safetensors import safe_open +from tqdm import tqdm +from einops import rearrange +from model.convert_from_ckpt import convert_ldm_unet_checkpoint, convert_ldm_clip_checkpoint, convert_ldm_vae_checkpoint +# from animatediff.utils.convert_lora_safetensor_to_diffusers import convert_lora, convert_motion_lora_ckpt_to_diffusers \ No newline at end of file diff --git a/modules/xadapter/xadapter_hijacks.py b/modules/xadapter/xadapter_hijacks.py new file mode 100644 index 000000000..2d1c1d327 --- /dev/null +++ b/modules/xadapter/xadapter_hijacks.py @@ -0,0 +1,58 @@ +import torch +from torch import nn + + +class FourierEmbedder(nn.Module): + def __init__(self, num_freqs=64, temperature=100): + super().__init__() + + self.num_freqs = num_freqs + self.temperature = temperature + + freq_bands = temperature ** (torch.arange(num_freqs) / num_freqs) + freq_bands = freq_bands[None, None, None] + self.register_buffer("freq_bands", freq_bands, persistent=False) + + def __call__(self, x): + x = self.freq_bands * x.unsqueeze(-1) + return torch.stack((x.sin(), x.cos()), dim=-1).permute(0, 1, 3, 4, 2).reshape(*x.shape[:2], -1) + + +class PositionNet(nn.Module): + def __init__(self, positive_len, out_dim, fourier_freqs=8): + super().__init__() + self.positive_len = positive_len + self.out_dim = out_dim + + self.fourier_embedder = FourierEmbedder(num_freqs=fourier_freqs) + self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy + + if isinstance(out_dim, tuple): + out_dim = out_dim[0] + self.linears = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + + self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim])) + + def forward(self, boxes, masks, positive_embeddings): + masks = masks.unsqueeze(-1) + + # embedding position (it may includes padding as placeholder) + xyxy_embedding = self.fourier_embedder(boxes) # B*N*4 -> B*N*C + + # learnable null embedding + positive_null = self.null_positive_feature.view(1, 1, -1) + xyxy_null = self.null_position_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null + xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null + + objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1)) + return objs diff --git a/pyproject.toml b/pyproject.toml index 699222d20..20b085569 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -15,6 +15,7 @@ exclude = [ "modules/control/units/*_model.py", "modules/control/units/*_pipe.py", "modules/pipelines/*.py", + "modules/xadapter/*.py", ] [tool.ruff.lint] select = [ diff --git a/scripts/init_latents.py b/scripts/init_latents.py index 08a2db205..4b8b286fa 100644 --- a/scripts/init_latents.py +++ b/scripts/init_latents.py @@ -1,9 +1,7 @@ -# from PIL import Image -# import gradio as gr -from modules import scripts, processing, shared, devices -from modules.processing_helpers import slerp import torch from diffusers.utils.torch_utils import randn_tensor +from modules import scripts, processing, shared, devices +from modules.processing_helpers import slerp class Script(scripts.Script): @@ -21,11 +19,9 @@ class Script(scripts.Script): generator = [torch.Generator(generator_device).manual_seed(s) for s in p.seeds] shape = (len(generator), shared.sd_model.unet.config.in_channels, p.height // shared.sd_model.vae_scale_factor, p.width // shared.sd_model.vae_scale_factor) - latents = randn_tensor(shape, generator=generator, device=shared.sd_model._execution_device, - dtype=shared.sd_model.unet.dtype) + latents = randn_tensor(shape, generator=generator, device=shared.sd_model._execution_device, dtype=shared.sd_model.unet.dtype) # pylint: disable=protected-access var_generator = [torch.Generator(generator_device).manual_seed(ss) for ss in p.subseeds] - var_latents = randn_tensor(shape, generator=var_generator, device=shared.sd_model._execution_device, - dtype=shared.sd_model.unet.dtype) + var_latents = randn_tensor(shape, generator=var_generator, device=shared.sd_model._execution_device, dtype=shared.sd_model.unet.dtype) # pylint: disable=protected-access return latents, var_latents, generator, var_generator @staticmethod @@ -44,9 +40,6 @@ class Script(scripts.Script): if shared.backend != shared.Backend.DIFFUSERS: return args = list(args) - if p.subseed_strength != 0: + if p.subseed_strength != 0 and getattr(shared.sd_model, '_execution_device', None) is not None: latents, var_latents, generator, var_generator = self.get_latents(p) self.set_slerp(p, latents, var_latents, generator, var_generator) - - - diff --git a/scripts/x_adapter.py b/scripts/x_adapter.py new file mode 100644 index 000000000..f6ea70db2 --- /dev/null +++ b/scripts/x_adapter.py @@ -0,0 +1,147 @@ +# https://github.com/showlab/X-Adapter + +import torch +import diffusers +import gradio as gr +import huggingface_hub as hf +from modules import errors, shared, devices, scripts, processing, sd_models, sd_samplers + + +adapter = None + + +class Script(scripts.Script): + def title(self): + return 'X-Adapter' + + def show(self, is_img2img): + return False + # return True if shared.backend == shared.Backend.DIFFUSERS else False + + def ui(self, _is_img2img): + with gr.Row(): + gr.HTML('  X-Adapter
') + with gr.Row(): + model = gr.Dropdown(label='Adapter model', choices=['None'] + sd_models.checkpoint_tiles(), value='None') + sampler = gr.Dropdown(label='Adapter sampler', choices=[s.name for s in sd_samplers.samplers], value='Default') + with gr.Row(): + width = gr.Slider(label='Adapter width', minimum=64, maximum=2048, step=8, value=512) + height = gr.Slider(label='Adapter height', minimum=64, maximum=2048, step=8, value=512) + with gr.Row(): + start = gr.Slider(label='Adapter start', minimum=0.0, maximum=1.0, step=0.01, value=0.5) + scale = gr.Slider(label='Adapter scale', minimum=0.0, maximum=1.0, step=0.01, value=1.0) + with gr.Row(): + lora = gr.Textbox('', label='Adapter LoRA', default='') + return model, sampler, width, height, start, scale, lora + + def run(self, p: processing.StableDiffusionProcessing, model, sampler, width, height, start, scale, lora): # pylint: disable=arguments-differ + from modules.xadapter.xadapter_hijacks import PositionNet + diffusers.models.embeddings.PositionNet = PositionNet # patch diffusers==0.26 from diffusers==0.20 + from modules.xadapter.adapter import Adapter_XL + from modules.xadapter.pipeline_sd_xl_adapter import StableDiffusionXLAdapterPipeline + from modules.xadapter.unet_adapter import UNet2DConditionModel as UNet2DConditionModelAdapter + + global adapter # pylint: disable=global-statement + if model == 'None': + return + else: + shared.opts.sd_model_refiner = model + if shared.sd_model_type != 'sdxl': + shared.log.error(f'X-Adapter: incorrect base model: {shared.sd_model.__class__.__name__}') + return + + if adapter is None: + shared.log.debug('X-Adapter: adapter loading') + adapter = Adapter_XL() + adapter_path = hf.hf_hub_download(repo_id='Lingmin-Ran/X-Adapter', filename='X_Adapter_v1.bin') + adapter_dict = torch.load(adapter_path) + adapter.load_state_dict(adapter_dict) + try: + if adapter is not None: + adapter.to(devices.device) + except Exception: + pass + if adapter is None: + shared.log.error('X-Adapter: adapter loading failed') + return + + sd_models.unload_model_weights(op='model') + sd_models.unload_model_weights(op='refiner') + orig_unetcondmodel = diffusers.models.unets.unet_2d_condition.UNet2DConditionModel + diffusers.models.UNet2DConditionModel = UNet2DConditionModelAdapter # patch diffusers with x-adapter + diffusers.models.unets.unet_2d_condition.UNet2DConditionModel = UNet2DConditionModelAdapter # patch diffusers with x-adapter + sd_models.reload_model_weights(op='model') + sd_models.reload_model_weights(op='refiner') + diffusers.models.unets.unet_2d_condition.UNet2DConditionModel = orig_unetcondmodel # unpatch diffusers + diffusers.models.UNet2DConditionModel = orig_unetcondmodel # unpatch diffusers + + if shared.sd_refiner_type != 'sd': + shared.log.error(f'X-Adapter: incorrect adapter model: {shared.sd_model.__class__.__name__}') + return + + # backup pipeline and params + orig_pipeline = shared.sd_model + orig_prompt_attention = shared.opts.prompt_attention + pipe = None + + try: + shared.log.debug('X-Adapter: creating pipeline') + pipe = StableDiffusionXLAdapterPipeline( + vae=shared.sd_model.vae, + text_encoder=shared.sd_model.text_encoder, + text_encoder_2=shared.sd_model.text_encoder_2, + tokenizer=shared.sd_model.tokenizer, + tokenizer_2=shared.sd_model.tokenizer_2, + unet=shared.sd_model.unet, + scheduler=shared.sd_model.scheduler, + vae_sd1_5=shared.sd_refiner.vae, + text_encoder_sd1_5=shared.sd_refiner.text_encoder, + tokenizer_sd1_5=shared.sd_refiner.tokenizer, + unet_sd1_5=shared.sd_refiner.unet, + scheduler_sd1_5=shared.sd_refiner.scheduler, + adapter=adapter, + ) + sd_models.copy_diffuser_options(pipe, shared.sd_model) + sd_models.set_diffuser_options(pipe) + try: + pipe.to(device=devices.device, dtype=devices.dtype) + except Exception: + pass + shared.opts.data['prompt_attention'] = 'Fixed attention' + prompt = shared.prompt_styles.apply_styles_to_prompt(p.prompt, p.styles) + negative = shared.prompt_styles.apply_negative_styles_to_prompt(p.negative_prompt, p.styles) + p.task_args['prompt'] = prompt + p.task_args['negative_prompt'] = negative + p.task_args['prompt_sd1_5'] = prompt + p.task_args['width_sd1_5'] = width + p.task_args['height_sd1_5'] = height + p.task_args['adapter_guidance_start'] = start + p.task_args['adapter_condition_scale'] = scale + p.task_args['fusion_guidance_scale'] = 1.0 # ??? + if sampler != 'Default': + pipe.scheduler_sd1_5 = sd_samplers.create_sampler(sampler, shared.sd_refiner) + else: + pipe.scheduler = diffusers.DPMSolverMultistepScheduler.from_config(pipe.scheduler.config) + pipe.scheduler_sd1_5 = diffusers.DPMSolverMultistepScheduler.from_config(pipe.scheduler_sd1_5.config) + pipe.scheduler_sd1_5.config.timestep_spacing = "leading" + shared.log.debug(f'X-Adapter: pipeline={pipe.__class__.__name__} args={p.task_args}') + shared.sd_model = pipe + except Exception as e: + shared.log.error(f'X-Adapter: pipeline creation failed: {e}') + errors.display(e, 'X-Adapter: pipeline creation failed') + shared.sd_model = orig_pipeline + + # run pipeline + processed: processing.Processed = processing.process_images(p) # runs processing using main loop + + # restore pipeline and params + try: + if adapter is not None: + adapter.to(devices.cpu) + except Exception: + pass + pipe = None + shared.opts.data['prompt_attention'] = orig_prompt_attention + shared.sd_model = orig_pipeline + devices.torch_gc() + return processed