mirror of
https://github.com/vladmandic/automatic
synced 2026-09-18 16:54:33 +02:00
Vladimir Mandic
666 lines
30 KiB
Python
666 lines
30 KiB
Python
# Copyright 2025 Microsoft and 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
|
|
from typing import Any, Callable
|
|
|
|
import numpy as np
|
|
import torch
|
|
from transformers import Qwen2Tokenizer, Qwen3VLForConditionalGeneration
|
|
|
|
from diffusers.image_processor import VaeImageProcessor
|
|
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
|
from diffusers.utils import is_torch_xla_available, logging, replace_example_docstring
|
|
from diffusers.utils.torch_utils import randn_tensor
|
|
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
|
from .transformer_mage_flow import MageFlowTransformer2DModel
|
|
from .pipeline_output import MageFlowPipelineOutput
|
|
from .autoencoder_mage_vae import AutoencoderMageVAE
|
|
|
|
|
|
if is_torch_xla_available():
|
|
import torch_xla.core.xla_model as xm
|
|
|
|
XLA_AVAILABLE = True
|
|
else:
|
|
XLA_AVAILABLE = False
|
|
|
|
|
|
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
|
|
|
EXAMPLE_DOC_STRING = """
|
|
Examples:
|
|
```py
|
|
>>> import torch
|
|
>>> from diffusers import MageFlowPipeline
|
|
|
|
>>> pipe = MageFlowPipeline.from_pretrained("microsoft/Mage-Flow-4B", torch_dtype=torch.bfloat16)
|
|
>>> pipe.to("cuda")
|
|
>>> prompt = "A cat holding a sign that says hello world"
|
|
>>> image = pipe(prompt, num_inference_steps=30, guidance_scale=5.0).images[0]
|
|
>>> image.save("mage_flow.png")
|
|
```
|
|
"""
|
|
|
|
|
|
# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
|
|
def retrieve_timesteps(
|
|
scheduler,
|
|
num_inference_steps: int | None = None,
|
|
device: str | torch.device | None = None,
|
|
timesteps: list[int] | None = None,
|
|
sigmas: list[float] | None = None,
|
|
**kwargs,
|
|
):
|
|
r"""
|
|
Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
|
|
custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
|
|
|
|
Args:
|
|
scheduler (`SchedulerMixin`):
|
|
The scheduler to get timesteps from.
|
|
num_inference_steps (`int`):
|
|
The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
|
|
must be `None`.
|
|
device (`str` or `torch.device`, *optional*):
|
|
The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
|
|
timesteps (`list[int]`, *optional*):
|
|
Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
|
|
`num_inference_steps` and `sigmas` must be `None`.
|
|
sigmas (`list[float]`, *optional*):
|
|
Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
|
|
`num_inference_steps` and `timesteps` must be `None`.
|
|
|
|
Returns:
|
|
`tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
|
|
second element is the number of inference steps.
|
|
"""
|
|
if timesteps is not None and sigmas is not None:
|
|
raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
|
|
if timesteps is not None:
|
|
accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accepts_timesteps:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" timestep schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
elif sigmas is not None:
|
|
accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
|
|
if not accept_sigmas:
|
|
raise ValueError(
|
|
f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
|
|
f" sigmas schedules. Please check whether you are using the correct scheduler."
|
|
)
|
|
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
num_inference_steps = len(timesteps)
|
|
else:
|
|
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
|
timesteps = scheduler.timesteps
|
|
return timesteps, num_inference_steps
|
|
|
|
|
|
class MageFlowPipeline(DiffusionPipeline):
|
|
r"""
|
|
The Mage-Flow pipeline for text-to-image generation.
|
|
|
|
Args:
|
|
transformer ([`MageFlowTransformer2DModel`]):
|
|
Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
|
|
scheduler ([`FlowMatchEulerDiscreteScheduler`]):
|
|
A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
|
|
vae ([`AutoencoderMageVAE`]):
|
|
Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
|
|
text_encoder ([`Qwen3VLForConditionalGeneration`]):
|
|
Qwen3-VL text encoder for producing text conditioning embeddings.
|
|
tokenizer (`AutoTokenizer`):
|
|
Tokenizer for the Qwen3-VL text encoder.
|
|
"""
|
|
|
|
model_cpu_offload_seq = "text_encoder->transformer->vae"
|
|
_callback_tensor_inputs = ["latents", "prompt_embeds"]
|
|
|
|
def __init__(
|
|
self,
|
|
scheduler: FlowMatchEulerDiscreteScheduler,
|
|
vae: AutoencoderMageVAE,
|
|
text_encoder: Qwen3VLForConditionalGeneration,
|
|
tokenizer: Qwen2Tokenizer,
|
|
transformer: MageFlowTransformer2DModel,
|
|
):
|
|
super().__init__()
|
|
|
|
self.register_modules(
|
|
vae=vae,
|
|
text_encoder=text_encoder,
|
|
tokenizer=tokenizer,
|
|
transformer=transformer,
|
|
scheduler=scheduler,
|
|
)
|
|
self.vae_scale_factor = 16 # MageVAE downsample factor
|
|
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
|
|
self.tokenizer_max_length = 2048
|
|
self.default_sample_size = 64 # 1024 / 16 = 64
|
|
# ChatML prompt template (same as QwenImage)
|
|
self.prompt_template = (
|
|
"<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, "
|
|
"text, spatial relationships of the objects and background:"
|
|
"<|im_end|>\n<|im_start|>user\n{}<|im_end|>\n<|im_start|>assistant\n"
|
|
)
|
|
self.prompt_template_start_idx = 34 # number of system-prompt tokens to skip
|
|
|
|
def _get_prompt_embeds(
|
|
self,
|
|
prompt: str | list[str] | None = None,
|
|
device: torch.device | None = None,
|
|
dtype: torch.dtype | None = None,
|
|
):
|
|
device = device or self._execution_device
|
|
if self.text_encoder is None:
|
|
raise ValueError(
|
|
"Text encoder is not available. Please provide `prompt_embeds` directly "
|
|
"when the pipeline is initialized without a text encoder."
|
|
)
|
|
dtype = dtype or self.text_encoder.dtype
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
|
|
template = self.prompt_template
|
|
drop_idx = self.prompt_template_start_idx
|
|
txt = [template.format(e) for e in prompt]
|
|
txt_tokens = self.tokenizer(
|
|
txt, max_length=self.tokenizer_max_length + drop_idx, padding=True, truncation=True, return_tensors="pt"
|
|
).to(device)
|
|
encoder_out = self.text_encoder(
|
|
input_ids=txt_tokens.input_ids,
|
|
attention_mask=txt_tokens.attention_mask,
|
|
output_hidden_states=True,
|
|
)
|
|
hidden_states = encoder_out.hidden_states[-1]
|
|
|
|
# Extract valid tokens per sample, drop system prompt prefix, then re-pad to uniform length.
|
|
bool_mask = txt_tokens.attention_mask.bool()
|
|
valid_lengths = bool_mask.sum(dim=1)
|
|
selected = hidden_states[bool_mask]
|
|
split_hidden_states = torch.split(selected, valid_lengths.tolist(), dim=0)
|
|
split_hidden_states = [e[drop_idx:] for e in split_hidden_states]
|
|
|
|
attn_mask_list = [torch.ones(e.size(0), dtype=torch.long, device=e.device) for e in split_hidden_states]
|
|
max_seq_len = max([e.size(0) for e in split_hidden_states])
|
|
prompt_embeds = torch.stack(
|
|
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0), u.size(1))]) for u in split_hidden_states]
|
|
)
|
|
prompt_embeds_mask = torch.stack(
|
|
[torch.cat([u, u.new_zeros(max_seq_len - u.size(0))]) for u in attn_mask_list]
|
|
)
|
|
|
|
prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)
|
|
|
|
return prompt_embeds, prompt_embeds_mask
|
|
|
|
def encode_prompt(
|
|
self,
|
|
prompt: str | list[str],
|
|
device: torch.device | None = None,
|
|
num_images_per_prompt: int = 1,
|
|
prompt_embeds: torch.Tensor | None = None,
|
|
prompt_embeds_mask: torch.Tensor | None = None,
|
|
max_sequence_length: int = 2048,
|
|
):
|
|
r"""
|
|
Encode the text prompt into embeddings for the transformer.
|
|
|
|
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.
|
|
prompt_embeds (`torch.Tensor`, *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.
|
|
prompt_embeds_mask (`torch.Tensor`, *optional*):
|
|
Attention mask for `prompt_embeds`.
|
|
max_sequence_length (`int`):
|
|
Maximum sequence length for the text embeddings.
|
|
"""
|
|
device = device or self._execution_device
|
|
|
|
prompt = [prompt] if isinstance(prompt, str) else prompt
|
|
batch_size = len(prompt) if prompt_embeds is None else prompt_embeds.shape[0]
|
|
|
|
if prompt_embeds is None:
|
|
prompt_embeds, prompt_embeds_mask = self._get_prompt_embeds(prompt, device)
|
|
|
|
prompt_embeds = prompt_embeds[:, :max_sequence_length]
|
|
_, seq_len, _ = prompt_embeds.shape
|
|
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
|
prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
|
|
|
if prompt_embeds_mask is not None:
|
|
prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]
|
|
prompt_embeds_mask = prompt_embeds_mask.repeat(1, num_images_per_prompt, 1)
|
|
prompt_embeds_mask = prompt_embeds_mask.view(batch_size * num_images_per_prompt, seq_len)
|
|
|
|
if prompt_embeds_mask.all():
|
|
prompt_embeds_mask = None
|
|
|
|
return prompt_embeds, prompt_embeds_mask
|
|
|
|
def check_inputs(
|
|
self,
|
|
prompt,
|
|
height,
|
|
width,
|
|
negative_prompt=None,
|
|
prompt_embeds=None,
|
|
negative_prompt_embeds=None,
|
|
prompt_embeds_mask=None,
|
|
negative_prompt_embeds_mask=None,
|
|
callback_on_step_end_tensor_inputs=None,
|
|
max_sequence_length=None,
|
|
):
|
|
if height % self.vae_scale_factor != 0 or width % self.vae_scale_factor != 0:
|
|
logger.warning(
|
|
f"`height` and `width` have to be divisible by {self.vae_scale_factor} but are {height} and {width}. "
|
|
"Dimensions will be resized accordingly"
|
|
)
|
|
|
|
if callback_on_step_end_tensor_inputs is not None and not all(
|
|
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
|
):
|
|
raise ValueError(
|
|
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found "
|
|
f"{[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
|
)
|
|
|
|
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 prompt_embeds_mask is None:
|
|
logger.warning(
|
|
"`prompt_embeds` is provided and `prompt_embeds_mask` is not provided, so the model will treat all"
|
|
" prompt tokens as valid. If `prompt_embeds` contains padding, you should provide the padding mask as"
|
|
" `prompt_embeds_mask`. Make sure to generate `prompt_embeds_mask` from the same text encoder that was"
|
|
" used to generate `prompt_embeds`."
|
|
)
|
|
|
|
if negative_prompt_embeds is not None and negative_prompt_embeds_mask is None:
|
|
logger.warning(
|
|
"`negative_prompt_embeds` is provided and `negative_prompt_embeds_mask` is not provided, so the model"
|
|
" will treat all negative prompt tokens as valid. If `negative_prompt_embeds` contains padding, you"
|
|
" should provide the padding mask as `negative_prompt_embeds_mask`. Make sure to generate"
|
|
" `negative_prompt_embeds_mask` from the same text encoder that was used to generate"
|
|
" `negative_prompt_embeds`."
|
|
)
|
|
|
|
if max_sequence_length is not None and max_sequence_length > 2048:
|
|
raise ValueError(f"`max_sequence_length` cannot be greater than 2048 but is {max_sequence_length}")
|
|
|
|
@staticmethod
|
|
def _prepare_latent_image_ids(height, width, device, dtype):
|
|
latent_image_ids = torch.zeros(height, width, 3, device=device, dtype=dtype)
|
|
latent_image_ids[..., 1] = torch.arange(height, device=device, dtype=dtype)[:, None]
|
|
latent_image_ids[..., 2] = torch.arange(width, device=device, dtype=dtype)[None, :]
|
|
latent_image_ids = latent_image_ids.reshape(height * width, 3)
|
|
return latent_image_ids
|
|
|
|
def prepare_latents(
|
|
self,
|
|
batch_size,
|
|
num_channels_latents,
|
|
height,
|
|
width,
|
|
dtype,
|
|
device,
|
|
generator,
|
|
latents=None,
|
|
):
|
|
# MageVAE: 16x downsample, no patch packing
|
|
height = height // self.vae_scale_factor
|
|
width = width // self.vae_scale_factor
|
|
|
|
shape = (batch_size, num_channels_latents, height, width)
|
|
|
|
if latents is not None:
|
|
return latents.to(device=device, dtype=dtype)
|
|
|
|
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."
|
|
)
|
|
|
|
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
|
# Flatten to sequence: [B, C, H, W] -> [B, H*W, C]
|
|
latents = latents.permute(0, 2, 3, 1).reshape(batch_size, height * width, num_channels_latents)
|
|
|
|
return latents
|
|
|
|
@property
|
|
def guidance_scale(self):
|
|
return self._guidance_scale
|
|
|
|
@property
|
|
def attention_kwargs(self):
|
|
return self._attention_kwargs
|
|
|
|
@property
|
|
def num_timesteps(self):
|
|
return self._num_timesteps
|
|
|
|
@property
|
|
def current_timestep(self):
|
|
return self._current_timestep
|
|
|
|
@property
|
|
def interrupt(self):
|
|
return self._interrupt
|
|
|
|
@torch.no_grad()
|
|
@replace_example_docstring(EXAMPLE_DOC_STRING)
|
|
def __call__(
|
|
self,
|
|
prompt: str | list[str] | None = None,
|
|
negative_prompt: str | list[str] | None = None,
|
|
height: int | None = None,
|
|
width: int | None = None,
|
|
num_inference_steps: int = 30,
|
|
guidance_scale: float = 5.0,
|
|
num_images_per_prompt: int = 1,
|
|
generator: torch.Generator | list[torch.Generator] | None = None,
|
|
latents: torch.Tensor | None = None,
|
|
prompt_embeds: torch.Tensor | None = None,
|
|
prompt_embeds_mask: torch.Tensor | None = None,
|
|
negative_prompt_embeds: torch.Tensor | None = None,
|
|
negative_prompt_embeds_mask: torch.Tensor | None = None,
|
|
output_type: str | None = "pil",
|
|
return_dict: bool = True,
|
|
attention_kwargs: dict[str, Any] | None = None,
|
|
callback_on_step_end: Callable[[int, int], None] | None = None,
|
|
callback_on_step_end_tensor_inputs: list[str] = ["latents"],
|
|
max_sequence_length: int = 2048,
|
|
sigmas: list[float] | None = None,
|
|
) -> MageFlowPipelineOutput | tuple:
|
|
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.
|
|
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
|
|
not greater than `1`).
|
|
height (`int`, *optional*, defaults to `self.default_sample_size * self.vae_scale_factor`):
|
|
The height in pixels of the generated image.
|
|
width (`int`, *optional*, defaults to `self.default_sample_size * self.vae_scale_factor`):
|
|
The width in pixels of the generated image.
|
|
num_inference_steps (`int`, *optional*, defaults to 30):
|
|
The number of denoising steps. More denoising steps usually lead to a higher quality image at the
|
|
expense of slower inference.
|
|
guidance_scale (`float`, *optional*, defaults to 5.0):
|
|
Classifier-free guidance scale. Enabled by setting `guidance_scale > 1`. Higher guidance scale
|
|
encourages images closely linked to the text `prompt`, usually at the expense of lower image quality.
|
|
num_images_per_prompt (`int`, *optional*, defaults to 1):
|
|
The number of images to generate per prompt.
|
|
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.Tensor`, *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 be generated by sampling using the supplied random `generator`.
|
|
prompt_embeds (`torch.Tensor`, *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.
|
|
prompt_embeds_mask (`torch.Tensor`, *optional*):
|
|
Attention mask for `prompt_embeds`.
|
|
negative_prompt_embeds (`torch.Tensor`, *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.
|
|
negative_prompt_embeds_mask (`torch.Tensor`, *optional*):
|
|
Attention mask for `negative_prompt_embeds`.
|
|
output_type (`str`, *optional*, defaults to `"pil"`):
|
|
The output format of the generated 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.mage_flow.MageFlowPipelineOutput`] instead of a plain tuple.
|
|
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).
|
|
callback_on_step_end (`Callable`, *optional*):
|
|
A function that calls at the end of each denoising steps during the inference. The function is called
|
|
with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
|
|
callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
|
|
`callback_on_step_end_tensor_inputs`.
|
|
callback_on_step_end_tensor_inputs (`list`, *optional*):
|
|
The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
|
|
will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
|
|
`._callback_tensor_inputs` attribute of your pipeline class.
|
|
max_sequence_length (`int`, defaults to 2048):
|
|
Maximum sequence length to use with the `prompt`.
|
|
sigmas (`list[float]`, *optional*):
|
|
Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
|
|
their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
|
|
will be used.
|
|
|
|
Examples:
|
|
|
|
Returns:
|
|
[`~pipelines.mage_flow.MageFlowPipelineOutput`] or `tuple`:
|
|
[`~pipelines.mage_flow.MageFlowPipelineOutput`] if `return_dict` is True, otherwise a `tuple`. When
|
|
returning a tuple, the first element is a list with the generated images.
|
|
"""
|
|
|
|
height = height or self.default_sample_size * self.vae_scale_factor
|
|
width = width or self.default_sample_size * self.vae_scale_factor
|
|
|
|
# 1. Check inputs. Raise error if not correct
|
|
self.check_inputs(
|
|
prompt,
|
|
height,
|
|
width,
|
|
negative_prompt=negative_prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
negative_prompt_embeds=negative_prompt_embeds,
|
|
prompt_embeds_mask=prompt_embeds_mask,
|
|
negative_prompt_embeds_mask=negative_prompt_embeds_mask,
|
|
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
|
max_sequence_length=max_sequence_length,
|
|
)
|
|
|
|
self._guidance_scale = guidance_scale
|
|
self._attention_kwargs = attention_kwargs
|
|
self._current_timestep = None
|
|
self._interrupt = False
|
|
|
|
# 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 = self._execution_device
|
|
|
|
do_classifier_free_guidance = guidance_scale > 1.0
|
|
|
|
# 3. Encode prompt
|
|
prompt_embeds, prompt_embeds_mask = self.encode_prompt(
|
|
prompt=prompt,
|
|
prompt_embeds=prompt_embeds,
|
|
prompt_embeds_mask=prompt_embeds_mask,
|
|
device=device,
|
|
num_images_per_prompt=num_images_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
)
|
|
if do_classifier_free_guidance:
|
|
if negative_prompt_embeds is None and self.text_encoder is None:
|
|
# text_encoder unavailable and no negative_prompt_embeds provided, skip CFG
|
|
do_classifier_free_guidance = False
|
|
else:
|
|
negative_prompt_embeds, negative_prompt_embeds_mask = self.encode_prompt(
|
|
prompt=negative_prompt if negative_prompt is not None else [""] * batch_size,
|
|
prompt_embeds=negative_prompt_embeds,
|
|
prompt_embeds_mask=negative_prompt_embeds_mask,
|
|
device=device,
|
|
num_images_per_prompt=num_images_per_prompt,
|
|
max_sequence_length=max_sequence_length,
|
|
)
|
|
|
|
# 4. Prepare latent variables
|
|
num_channels_latents = self.transformer.config.in_channels
|
|
latents = self.prepare_latents(
|
|
batch_size * num_images_per_prompt,
|
|
num_channels_latents,
|
|
height,
|
|
width,
|
|
prompt_embeds.dtype,
|
|
device,
|
|
generator,
|
|
latents,
|
|
)
|
|
|
|
# 5. Prepare image position ids for RoPE
|
|
latent_h = height // self.vae_scale_factor
|
|
latent_w = width // self.vae_scale_factor
|
|
img_ids = self._prepare_latent_image_ids(latent_h, latent_w, device, prompt_embeds.dtype)
|
|
|
|
# 6. Prepare timesteps
|
|
# Mage-Flow uses base_sigmas = linspace(1, 1/N, N) which differs from the scheduler's
|
|
# default sigma computation. The scheduler's shift=6.0 is applied on top of these.
|
|
if sigmas is None:
|
|
sigmas = np.linspace(1.0, 1.0 / num_inference_steps, num_inference_steps)
|
|
timesteps, num_inference_steps = retrieve_timesteps(
|
|
self.scheduler,
|
|
num_inference_steps,
|
|
device,
|
|
sigmas=sigmas,
|
|
)
|
|
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
|
self._num_timesteps = len(timesteps)
|
|
|
|
if self.attention_kwargs is None:
|
|
self._attention_kwargs = {}
|
|
|
|
# 7. Denoising loop
|
|
self.scheduler.set_begin_index(0)
|
|
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
|
for i, t in enumerate(timesteps):
|
|
if self.interrupt:
|
|
continue
|
|
|
|
self._current_timestep = t
|
|
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
|
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
|
|
|
if do_classifier_free_guidance:
|
|
noise_pred_cond = self.transformer(
|
|
hidden_states=latents,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep=timestep / 1000,
|
|
img_ids=img_ids,
|
|
joint_attention_kwargs=self.attention_kwargs,
|
|
return_dict=False,
|
|
)[0]
|
|
noise_pred_uncond = self.transformer(
|
|
hidden_states=latents,
|
|
encoder_hidden_states=negative_prompt_embeds,
|
|
timestep=timestep / 1000,
|
|
img_ids=img_ids,
|
|
joint_attention_kwargs=self.attention_kwargs,
|
|
return_dict=False,
|
|
)[0]
|
|
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
|
|
else:
|
|
noise_pred = self.transformer(
|
|
hidden_states=latents,
|
|
encoder_hidden_states=prompt_embeds,
|
|
timestep=timestep / 1000,
|
|
img_ids=img_ids,
|
|
joint_attention_kwargs=self.attention_kwargs,
|
|
return_dict=False,
|
|
)[0]
|
|
|
|
# compute the previous noisy sample x_t -> x_t-1
|
|
latents_dtype = latents.dtype
|
|
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
|
|
|
if latents.dtype != latents_dtype:
|
|
if torch.backends.mps.is_available():
|
|
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
|
latents = latents.to(latents_dtype)
|
|
|
|
if callback_on_step_end is not None:
|
|
callback_kwargs = {}
|
|
for k in callback_on_step_end_tensor_inputs:
|
|
callback_kwargs[k] = locals()[k]
|
|
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
|
|
|
latents = callback_outputs.pop("latents", latents)
|
|
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
|
|
|
# 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 XLA_AVAILABLE:
|
|
xm.mark_step()
|
|
|
|
self._current_timestep = None
|
|
|
|
# 8. VAE decode
|
|
if output_type == "latent":
|
|
image = latents
|
|
else:
|
|
# Unflatten: [B, H*W, C] -> [B, C, H, W]
|
|
latents = latents.reshape(batch_size * num_images_per_prompt, latent_h, latent_w, num_channels_latents)
|
|
latents = latents.permute(0, 3, 1, 2)
|
|
latents = latents.to(self.vae.dtype)
|
|
image = self.vae(latents, return_dict=False)[0]
|
|
image = image.clamp(-1, 1)
|
|
image = self.image_processor.postprocess(image, output_type=output_type)
|
|
|
|
# Offload all models
|
|
self.maybe_free_model_hooks()
|
|
|
|
if not return_dict:
|
|
return (image,)
|
|
|
|
return MageFlowPipelineOutput(images=image)
|