major refactoring of modules

Signed-off-by: Vladimir Mandic <mandic00@live.com>

scripts/instantir/__init__.py

@@ -0,0 +1,3 @@
+from .sdxl_instantir import InstantIRPipeline
+from .lcm_single_step_scheduler import LCMSingleStepScheduler
+from .ip_adapter.utils import init_adapter_in_unet, load_adapter_to_pipe

scripts/instantir/aggregator.py

@@ -0,0 +1,982 @@
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import BaseOutput, logging
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import TextImageProjection, TextImageTimeEmbedding, TextTimeEmbedding, TimestepEmbedding, Timesteps
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.unets.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    DownBlock2D,
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    get_down_block,
+)
+from diffusers.models.unets.unet_2d_condition import UNet2DConditionModel
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class ZeroConv(nn.Module):
+    def __init__(self, label_nc, norm_nc, mask=False):
+        super().__init__()
+        self.zero_conv = zero_module(nn.Conv2d(label_nc+norm_nc, norm_nc, 1, 1, 0))
+        self.mask = mask
+
+    def forward(self, hidden_states, h_ori=None):
+        # with torch.cuda.amp.autocast(enabled=False, dtype=torch.float32):
+        c, h = hidden_states
+        if not self.mask:
+            h = self.zero_conv(torch.cat([c, h], dim=1))
+        else:
+            h = self.zero_conv(torch.cat([c, h], dim=1)) * torch.zeros_like(h)
+        if h_ori is not None:
+            h = torch.cat([h_ori, h], dim=1)
+        return h
+
+
+class SFT(nn.Module):
+    def __init__(self, label_nc, norm_nc, mask=False):
+        super().__init__()
+
+        # param_free_norm_type = str(parsed.group(1))
+        ks = 3
+        pw = ks // 2
+
+        self.mask = mask
+
+        nhidden = 128
+
+        self.mlp_shared = nn.Sequential(
+            nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
+            nn.SiLU()
+        )
+        self.mul = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
+        self.add = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
+
+    def forward(self, hidden_states, mask=False):
+        
+        c, h = hidden_states
+        mask = mask or self.mask
+        assert mask is False
+
+        actv = self.mlp_shared(c)
+        gamma = self.mul(actv)
+        beta = self.add(actv)
+
+        if self.mask:
+            gamma = gamma * torch.zeros_like(gamma)
+            beta = beta * torch.zeros_like(beta)
+        # gamma_ori, gamma_res = torch.split(gamma, [h_ori_c, h_c], dim=1)
+        # beta_ori, beta_res = torch.split(beta, [h_ori_c, h_c], dim=1)
+        # print(gamma_ori.mean(), gamma_res.mean(), beta_ori.mean(), beta_res.mean())
+        h = h * (gamma + 1) + beta
+        # sample_ori, sample_res = torch.split(h, [h_ori_c, h_c], dim=1)
+        # print(sample_ori.mean(), sample_res.mean())
+
+        return h
+
+
+@dataclass
+class AggregatorOutput(BaseOutput):
+    """
+    The output of [`Aggregator`].
+
+    Args:
+        down_block_res_samples (`tuple[torch.Tensor]`):
+            A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
+            be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
+            used to condition the original UNet's downsampling activations.
+        mid_down_block_re_sample (`torch.Tensor`):
+            The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
+            `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
+            Output can be used to condition the original UNet's middle block activation.
+    """
+
+    down_block_res_samples: Tuple[torch.Tensor]
+    mid_block_res_sample: torch.Tensor
+
+
+class ConditioningEmbedding(nn.Module):
+    """
+    Quoting from https://arxiv.org/abs/2302.05543: "Stable Diffusion uses a pre-processing method similar to VQ-GAN
+    [11] to convert the entire dataset of 512 × 512 images into smaller 64 × 64 “latent images” for stabilized
+    training. This requires ControlNets to convert image-based conditions to 64 × 64 feature space to match the
+    convolution size. We use a tiny network E(·) of four convolution layers with 4 × 4 kernels and 2 × 2 strides
+    (activated by ReLU, channels are 16, 32, 64, 128, initialized with Gaussian weights, trained jointly with the full
+    model) to encode image-space conditions ... into feature maps ..."
+    """
+
+    def __init__(
+        self,
+        conditioning_embedding_channels: int,
+        conditioning_channels: int = 3,
+        block_out_channels: Tuple[int, ...] = (16, 32, 96, 256),
+    ):
+        super().__init__()
+
+        self.conv_in = nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+
+        for i in range(len(block_out_channels) - 1):
+            channel_in = block_out_channels[i]
+            channel_out = block_out_channels[i + 1]
+            self.blocks.append(nn.Conv2d(channel_in, channel_in, kernel_size=3, padding=1))
+            self.blocks.append(nn.Conv2d(channel_in, channel_out, kernel_size=3, padding=1, stride=2))
+
+        self.conv_out = zero_module(
+            nn.Conv2d(block_out_channels[-1], conditioning_embedding_channels, kernel_size=3, padding=1)
+        )
+
+    def forward(self, conditioning):
+        embedding = self.conv_in(conditioning)
+        embedding = F.silu(embedding)
+
+        for block in self.blocks:
+            embedding = block(embedding)
+            embedding = F.silu(embedding)
+
+        embedding = self.conv_out(embedding)
+
+        return embedding
+
+
+class Aggregator(ModelMixin, ConfigMixin, FromOriginalModelMixin):
+    """
+    Aggregator model.
+
+    Args:
+        in_channels (`int`, defaults to 4):
+            The number of channels in the input sample.
+        flip_sin_to_cos (`bool`, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, defaults to 0):
+            The frequency shift to apply to the time embedding.
+        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
+        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, defaults to 2):
+            The number of layers per block.
+        downsample_padding (`int`, defaults to 1):
+            The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, defaults to 1):
+            The scale factor to use for the mid block.
+        act_fn (`str`, 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`, defaults to 1e-5):
+            The epsilon to use for the normalization.
+        cross_attention_dim (`int`, 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 (`Union[int, Tuple[int]]`, defaults to 8):
+            The dimension of the attention heads.
+        use_linear_projection (`bool`, defaults to `False`):
+        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.
+        num_class_embeds (`int`, *optional*, defaults to 0):
+            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`.
+        upcast_attention (`bool`, defaults to `False`):
+        resnet_time_scale_shift (`str`, defaults to `"default"`):
+            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
+        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
+            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
+            `class_embed_type="projection"`.
+        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
+            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
+        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
+            The tuple of output channel for each block in the `conditioning_embedding` layer.
+        global_pool_conditions (`bool`, defaults to `False`):
+            TODO(Patrick) - unused parameter.
+        addition_embed_type_num_heads (`int`, defaults to 64):
+            The number of heads to use for the `TextTimeEmbedding` layer.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        in_channels: int = 4,
+        conditioning_channels: int = 3,
+        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",
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
+        layers_per_block: 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: 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,
+        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",
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        global_pool_conditions: bool = False,
+        addition_embed_type_num_heads: int = 64,
+        pad_concat: bool = False,
+    ):
+        super().__init__()
+
+        # 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
+        self.pad_concat = pad_concat
+
+        # Check inputs
+        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 isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        # input
+        conv_in_kernel = 3
+        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
+        time_embed_dim = 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]
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+        )
+
+        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"` (Kandinsky 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 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)
+        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)
+        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"` (Kandinsky 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 is not None:
+            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")
+
+        # control net conditioning embedding
+        self.ref_conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        self.down_blocks = nn.ModuleList([])
+        self.controlnet_down_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        # down
+        output_channel = block_out_channels[0]
+
+        # controlnet_block = ZeroConv(output_channel, output_channel)
+        controlnet_block = nn.Sequential(
+            SFT(output_channel, output_channel),
+            zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+        )
+        self.controlnet_down_blocks.append(controlnet_block)
+
+        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,
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=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,
+                num_attention_heads=num_attention_heads[i],
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                downsample_padding=downsample_padding,
+                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,
+            )
+            self.down_blocks.append(down_block)
+
+            for _ in range(layers_per_block):
+                # controlnet_block = ZeroConv(output_channel, output_channel)
+                controlnet_block = nn.Sequential(
+                    SFT(output_channel, output_channel),
+                    zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+                )
+                self.controlnet_down_blocks.append(controlnet_block)
+
+            if not is_final_block:
+                # controlnet_block = ZeroConv(output_channel, output_channel)
+                controlnet_block = nn.Sequential(
+                    SFT(output_channel, output_channel),
+                    zero_module(nn.Conv2d(output_channel, output_channel, kernel_size=1))
+                )
+                self.controlnet_down_blocks.append(controlnet_block)
+
+        # mid
+        mid_block_channel = block_out_channels[-1]
+
+        # controlnet_block = ZeroConv(mid_block_channel, mid_block_channel)
+        controlnet_block = nn.Sequential(
+            SFT(mid_block_channel, mid_block_channel),
+            zero_module(nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1))
+        )
+        self.controlnet_mid_block = controlnet_block
+
+        if mid_block_type == "UNetMidBlock2DCrossAttn":
+            self.mid_block = UNetMidBlock2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=mid_block_channel,
+                temb_channels=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,
+                num_attention_heads=num_attention_heads[-1],
+                resnet_groups=norm_num_groups,
+                use_linear_projection=use_linear_projection,
+                upcast_attention=upcast_attention,
+            )
+        elif mid_block_type == "UNetMidBlock2D":
+            self.mid_block = UNetMidBlock2D(
+                in_channels=block_out_channels[-1],
+                temb_channels=time_embed_dim,
+                num_layers=0,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                output_scale_factor=mid_block_scale_factor,
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                add_attention=False,
+            )
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+
+    @classmethod
+    def from_unet(
+        cls,
+        unet: UNet2DConditionModel,
+        controlnet_conditioning_channel_order: str = "rgb",
+        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+        load_weights_from_unet: bool = True,
+        conditioning_channels: int = 3,
+    ):
+        r"""
+        Instantiate a [`ControlNetModel`] from [`UNet2DConditionModel`].
+
+        Parameters:
+            unet (`UNet2DConditionModel`):
+                The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
+                where applicable.
+        """
+        transformer_layers_per_block = (
+            unet.config.transformer_layers_per_block if "transformer_layers_per_block" in unet.config else 1
+        )
+        encoder_hid_dim = unet.config.encoder_hid_dim if "encoder_hid_dim" in unet.config else None
+        encoder_hid_dim_type = unet.config.encoder_hid_dim_type if "encoder_hid_dim_type" in unet.config else None
+        addition_embed_type = unet.config.addition_embed_type if "addition_embed_type" in unet.config else None
+        addition_time_embed_dim = (
+            unet.config.addition_time_embed_dim if "addition_time_embed_dim" in unet.config else None
+        )
+
+        controlnet = cls(
+            encoder_hid_dim=encoder_hid_dim,
+            encoder_hid_dim_type=encoder_hid_dim_type,
+            addition_embed_type=addition_embed_type,
+            addition_time_embed_dim=addition_time_embed_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=unet.config.in_channels,
+            flip_sin_to_cos=unet.config.flip_sin_to_cos,
+            freq_shift=unet.config.freq_shift,
+            down_block_types=unet.config.down_block_types,
+            only_cross_attention=unet.config.only_cross_attention,
+            block_out_channels=unet.config.block_out_channels,
+            layers_per_block=unet.config.layers_per_block,
+            downsample_padding=unet.config.downsample_padding,
+            mid_block_scale_factor=unet.config.mid_block_scale_factor,
+            act_fn=unet.config.act_fn,
+            norm_num_groups=unet.config.norm_num_groups,
+            norm_eps=unet.config.norm_eps,
+            cross_attention_dim=unet.config.cross_attention_dim,
+            attention_head_dim=unet.config.attention_head_dim,
+            num_attention_heads=unet.config.num_attention_heads,
+            use_linear_projection=unet.config.use_linear_projection,
+            class_embed_type=unet.config.class_embed_type,
+            num_class_embeds=unet.config.num_class_embeds,
+            upcast_attention=unet.config.upcast_attention,
+            resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+            projection_class_embeddings_input_dim=unet.config.projection_class_embeddings_input_dim,
+            mid_block_type=unet.config.mid_block_type,
+            controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+            conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+            conditioning_channels=conditioning_channels,
+        )
+
+        if load_weights_from_unet:
+            controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.ref_conv_in.load_state_dict(unet.conv_in.state_dict())
+            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
+            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())
+
+            if controlnet.class_embedding:
+                controlnet.class_embedding.load_state_dict(unet.class_embedding.state_dict())
+
+            if hasattr(controlnet, "add_embedding"):
+                controlnet.add_embedding.load_state_dict(unet.add_embedding.state_dict())
+
+            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
+            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())
+
+        return controlnet
+
+    @property
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
+    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, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True)
+
+            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
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
+    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)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attention_slice
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]]) -> None:
+        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 process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        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":
+            # Kandinsky 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)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+    
+    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+        if isinstance(module, (CrossAttnDownBlock2D, DownBlock2D)):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        controlnet_cond: torch.FloatTensor,
+        cat_dim: int = -2,
+        conditioning_scale: float = 1.0,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        return_dict: bool = True,
+    ) -> Union[AggregatorOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
+        """
+        The [`Aggregator`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor.
+            timestep (`Union[torch.Tensor, float, int]`):
+                The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states.
+            controlnet_cond (`torch.FloatTensor`):
+                The conditional input tensor of shape `(batch_size, sequence_length, hidden_size)`.
+            conditioning_scale (`float`, defaults to `1.0`):
+                The scale factor for ControlNet outputs.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
+                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
+                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
+                embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            added_cond_kwargs (`dict`):
+                Additional conditions for the Stable Diffusion XL UNet.
+            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
+            return_dict (`bool`, defaults to `True`):
+                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.
+
+        Returns:
+            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
+                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
+                returned where the first element is the sample tensor.
+        """
+        # check channel order
+        channel_order = self.config.controlnet_conditioning_channel_order
+
+        if channel_order == "rgb":
+            # in rgb order by default
+            ...
+        else:
+            raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")
+
+        # prepare attention_mask
+        if attention_mask is not None:
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # 1. time
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # 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)
+
+        # 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)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=self.dtype)
+            emb = emb + class_emb
+
+        if self.config.addition_embed_type is not None:
+            if self.config.addition_embed_type == "text":
+                aug_emb = self.add_embedding(encoder_hidden_states)
+
+            elif self.config.addition_embed_type == "text_time":
+                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)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. prepare input
+        cond_latent = self.conv_in(sample)
+        ref_latent = self.ref_conv_in(controlnet_cond)
+        batch_size, channel, height, width = cond_latent.shape
+        if self.pad_concat:
+            if cat_dim == -2 or cat_dim == 2:
+                concat_pad = torch.zeros(batch_size, channel, 1, width)
+            elif cat_dim == -1 or cat_dim == 3:
+                concat_pad = torch.zeros(batch_size, channel, height, 1)
+            else:
+                raise ValueError(f"Aggregator shall concat along spatial dimension, but is asked to concat dim: {cat_dim}.")
+            concat_pad = concat_pad.to(cond_latent.device, dtype=cond_latent.dtype)
+            sample = torch.cat([cond_latent, concat_pad, ref_latent], dim=cat_dim)
+        else:
+            sample = torch.cat([cond_latent, ref_latent], dim=cat_dim)
+
+        # 3. down
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            sample, res_samples = downsample_block(
+                hidden_states=sample,
+                temb=emb,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+            # rebuild sample: split and concat
+            if self.pad_concat:
+                batch_size, channel, height, width = sample.shape
+                if cat_dim == -2 or cat_dim == 2:
+                    cond_latent = sample[:, :, :height//2, :]
+                    ref_latent = sample[:, :, -(height//2):, :]
+                    concat_pad = torch.zeros(batch_size, channel, 1, width)
+                elif cat_dim == -1 or cat_dim == 3:
+                    cond_latent = sample[:, :, :, :width//2]
+                    ref_latent = sample[:, :, :, -(width//2):]
+                    concat_pad = torch.zeros(batch_size, channel, height, 1)
+                concat_pad = concat_pad.to(cond_latent.device, dtype=cond_latent.dtype)
+                sample = torch.cat([cond_latent, concat_pad, ref_latent], dim=cat_dim)
+                res_samples = res_samples[:-1] + (sample,)
+
+            down_block_res_samples += res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                cross_attention_kwargs=cross_attention_kwargs,
+            )
+
+        # 5. split samples and SFT.
+        controlnet_down_block_res_samples = ()
+        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+            batch_size, channel, height, width = down_block_res_sample.shape
+            if cat_dim == -2 or cat_dim == 2:
+                cond_latent = down_block_res_sample[:, :, :height//2, :]
+                ref_latent = down_block_res_sample[:, :, -(height//2):, :]
+            elif cat_dim == -1 or cat_dim == 3:
+                cond_latent = down_block_res_sample[:, :, :, :width//2]
+                ref_latent = down_block_res_sample[:, :, :, -(width//2):]
+            down_block_res_sample = controlnet_block((cond_latent, ref_latent), )
+            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+        down_block_res_samples = controlnet_down_block_res_samples
+
+        batch_size, channel, height, width = sample.shape
+        if cat_dim == -2 or cat_dim == 2:
+            cond_latent = sample[:, :, :height//2, :]
+            ref_latent = sample[:, :, -(height//2):, :]
+        elif cat_dim == -1 or cat_dim == 3:
+            cond_latent = sample[:, :, :, :width//2]
+            ref_latent = sample[:, :, :, -(width//2):]
+        mid_block_res_sample = self.controlnet_mid_block((cond_latent, ref_latent), )
+
+        # 6. scaling
+        down_block_res_samples = [sample*conditioning_scale for sample in down_block_res_samples]
+        mid_block_res_sample = mid_block_res_sample*conditioning_scale
+
+        if self.config.global_pool_conditions:
+            down_block_res_samples = [
+                torch.mean(sample, dim=(2, 3), keepdim=True) for sample in down_block_res_samples
+            ]
+            mid_block_res_sample = torch.mean(mid_block_res_sample, dim=(2, 3), keepdim=True)
+
+        if not return_dict:
+            return (down_block_res_samples, mid_block_res_sample)
+
+        return AggregatorOutput(
+            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+        )
+
+
+def zero_module(module):
+    for p in module.parameters():
+        nn.init.zeros_(p)
+    return module

scripts/instantir/ip_adapter/ip_adapter.py

@@ -0,0 +1,234 @@
+import os
+import torch
+from typing import List
+from collections import namedtuple, OrderedDict
+
+def is_torch2_available():
+    return hasattr(torch.nn.functional, "scaled_dot_product_attention")
+
+if is_torch2_available():
+    from .attention_processor import (
+        AttnProcessor2_0 as AttnProcessor,
+    )
+    from .attention_processor import (
+        CNAttnProcessor2_0 as CNAttnProcessor,
+    )
+    from .attention_processor import (
+        IPAttnProcessor2_0 as IPAttnProcessor,
+    )
+    from .attention_processor import (
+        TA_IPAttnProcessor2_0 as TA_IPAttnProcessor,
+    )
+else:
+    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor, TA_IPAttnProcessor
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=2048, clip_embeddings_dim=1280, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+    def __init__(self, cross_attention_dim=2048, clip_embeddings_dim=1280):
+        super().__init__()
+        
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim)
+        )
+        
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class MultiIPAdapterImageProjection(torch.nn.Module):
+    def __init__(self, IPAdapterImageProjectionLayers):
+        super().__init__()
+        self.image_projection_layers = torch.nn.ModuleList(IPAdapterImageProjectionLayers)
+
+    def forward(self, image_embeds: List[torch.FloatTensor]):
+        projected_image_embeds = []
+
+        # currently, we accept `image_embeds` as
+        #  1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim]
+        #  2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim]
+        if not isinstance(image_embeds, list):
+            image_embeds = [image_embeds.unsqueeze(1)]
+
+        if len(image_embeds) != len(self.image_projection_layers):
+            raise ValueError(
+                f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}"
+            )
+
+        for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers):
+            batch_size, num_images = image_embed.shape[0], image_embed.shape[1]
+            image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:])
+            image_embed = image_projection_layer(image_embed)
+            # image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:])
+
+            projected_image_embeds.append(image_embed)
+
+        return projected_image_embeds
+
+
+class IPAdapter(torch.nn.Module):
+    """IP-Adapter"""
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj = image_proj_model
+        self.ip_adapter = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+        keys = list(state_dict.keys())
+        if keys != ["image_proj", "ip_adapter"]:
+            state_dict = revise_state_dict(state_dict)
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj.load_state_dict(state_dict["image_proj"], strict=True)
+        self.ip_adapter.load_state_dict(state_dict["ip_adapter"], strict=True)
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_modules did not change!"
+
+
+class IPAdapterPlus(torch.nn.Module):
+    """IP-Adapter"""
+    def __init__(self, unet, image_proj_model, adapter_modules, ckpt_path=None):
+        super().__init__()
+        self.unet = unet
+        self.image_proj = image_proj_model
+        self.ip_adapter = adapter_modules
+
+        if ckpt_path is not None:
+            self.load_from_checkpoint(ckpt_path)
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
+        return noise_pred
+
+    def load_from_checkpoint(self, ckpt_path: str):
+        # Calculate original checksums
+        orig_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        orig_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+        org_unet_sum = []
+        for attn_name, attn_proc in self.unet.attn_processors.items():
+            if isinstance(attn_proc, (TA_IPAttnProcessor, IPAttnProcessor)):
+                org_unet_sum.append(torch.sum(torch.stack([torch.sum(p) for p in attn_proc.parameters()])))
+        org_unet_sum = torch.sum(torch.stack(org_unet_sum))
+
+        state_dict = torch.load(ckpt_path, map_location="cpu")
+        keys = list(state_dict.keys())
+        if keys != ["image_proj", "ip_adapter"]:
+            state_dict = revise_state_dict(state_dict)
+
+        # Check if 'latents' exists in both the saved state_dict and the current model's state_dict
+        strict_load_image_proj_model = True
+        if "latents" in state_dict["image_proj"] and "latents" in self.image_proj.state_dict():
+            # Check if the shapes are mismatched
+            if state_dict["image_proj"]["latents"].shape != self.image_proj.state_dict()["latents"].shape:
+                del state_dict["image_proj"]["latents"]
+                strict_load_image_proj_model = False
+
+        # Load state dict for image_proj_model and adapter_modules
+        self.image_proj.load_state_dict(state_dict["image_proj"], strict=strict_load_image_proj_model)
+        missing_key, unexpected_key = self.ip_adapter.load_state_dict(state_dict["ip_adapter"], strict=False)
+        if len(missing_key) > 0:
+            for ms in missing_key:
+                if "ln" not in ms:
+                    raise ValueError(f"Missing key in adapter_modules: {len(missing_key)}")
+        if len(unexpected_key) > 0:
+            raise ValueError(f"Unexpected key in adapter_modules: {len(unexpected_key)}")
+
+        # Calculate new checksums
+        new_ip_proj_sum = torch.sum(torch.stack([torch.sum(p) for p in self.image_proj.parameters()]))
+        new_adapter_sum = torch.sum(torch.stack([torch.sum(p) for p in self.ip_adapter.parameters()]))
+
+        # Verify if the weights loaded to unet
+        unet_sum = []
+        for attn_name, attn_proc in self.unet.attn_processors.items():
+            if isinstance(attn_proc, (TA_IPAttnProcessor, IPAttnProcessor)):
+                unet_sum.append(torch.sum(torch.stack([torch.sum(p) for p in attn_proc.parameters()])))
+        unet_sum = torch.sum(torch.stack(unet_sum))
+
+        assert org_unet_sum != unet_sum, "Weights of adapter_modules in unet did not change!"
+        assert (unet_sum - new_adapter_sum < 1e-4), "Weights of adapter_modules did not load to unet!"
+
+        # Verify if the weights have changed
+        assert orig_ip_proj_sum != new_ip_proj_sum, "Weights of image_proj_model did not change!"
+        assert orig_adapter_sum != new_adapter_sum, "Weights of adapter_mod`ules did not change!"
+
+
+class IPAdapterXL(IPAdapter):
+    """SDXL"""
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, unet_added_cond_kwargs, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states, added_cond_kwargs=unet_added_cond_kwargs).sample
+        return noise_pred
+
+
+class IPAdapterPlusXL(IPAdapterPlus):
+    """IP-Adapter with fine-grained features"""
+
+    def forward(self, noisy_latents, timesteps, encoder_hidden_states, unet_added_cond_kwargs, image_embeds):
+        ip_tokens = self.image_proj(image_embeds)
+        encoder_hidden_states = torch.cat([encoder_hidden_states, ip_tokens], dim=1)
+        # Predict the noise residual
+        noise_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states, added_cond_kwargs=unet_added_cond_kwargs).sample
+        return noise_pred
+
+
+class IPAdapterFull(IPAdapterPlus):
+    """IP-Adapter with full features"""
+
+    def init_proj(self):
+        image_proj_model = MLPProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.hidden_size,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model

scripts/instantir/ip_adapter/resampler.py

@@ -0,0 +1,158 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1280,
+        depth=4,
+        dim_head=64,
+        heads=20,
+        num_queries=64,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)

scripts/instantir/ip_adapter/utils.py

@@ -0,0 +1,248 @@
+import torch
+from collections import namedtuple, OrderedDict
+from safetensors import safe_open
+from .attention_processor import init_attn_proc
+from .ip_adapter import MultiIPAdapterImageProjection
+from .resampler import Resampler
+from transformers import (
+    AutoModel, AutoImageProcessor,
+    CLIPVisionModelWithProjection, CLIPImageProcessor)
+
+
+def init_adapter_in_unet(
+        unet,
+        image_proj_model=None,
+        pretrained_model_path_or_dict=None,
+        adapter_tokens=64,
+        embedding_dim=None,
+        use_lcm=False,
+        use_adaln=True,
+    ):
+        device = unet.device
+        dtype = unet.dtype
+        if image_proj_model is None:
+            assert embedding_dim is not None, "embedding_dim must be provided if image_proj_model is None."
+            image_proj_model = Resampler(
+                embedding_dim=embedding_dim,
+                output_dim=unet.config.cross_attention_dim,
+                num_queries=adapter_tokens,
+            )
+        if pretrained_model_path_or_dict is not None:
+            if not isinstance(pretrained_model_path_or_dict, dict):
+                if pretrained_model_path_or_dict.endswith(".safetensors"):
+                    state_dict = {"image_proj": {}, "ip_adapter": {}}
+                    with safe_open(pretrained_model_path_or_dict, framework="pt", device=unet.device) as f:
+                        for key in f.keys():
+                            if key.startswith("image_proj."):
+                                state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                            elif key.startswith("ip_adapter."):
+                                state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+                else:
+                    state_dict = torch.load(pretrained_model_path_or_dict, map_location=unet.device)
+            else:
+                state_dict = pretrained_model_path_or_dict
+            keys = list(state_dict.keys())
+            if "image_proj" not in keys and "ip_adapter" not in keys:
+                state_dict = revise_state_dict(state_dict)
+
+        # Creat IP cross-attention in unet.
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
+        unet.set_attn_processor(attn_procs)
+
+        # Load pretrinaed model if needed.
+        if pretrained_model_path_or_dict is not None:
+            if "ip_adapter" in state_dict.keys():
+                adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+                missing, unexpected = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
+                for mk in missing:
+                    if "ln" not in mk:
+                        raise ValueError(f"Missing keys in adapter_modules: {missing}")
+            if "image_proj" in state_dict.keys():
+                image_proj_model.load_state_dict(state_dict["image_proj"])
+
+        # Load image projectors into iterable ModuleList.
+        image_projection_layers = []
+        image_projection_layers.append(image_proj_model)
+        unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+
+        # Adjust unet config to handle addtional ip hidden states.
+        unet.config.encoder_hid_dim_type = "ip_image_proj"
+        unet.to(dtype=dtype, device=device)
+
+
+def load_adapter_to_pipe(
+        pipe,
+        pretrained_model_path_or_dict,
+        image_encoder_or_path=None,
+        feature_extractor_or_path=None,
+        use_clip_encoder=False,
+        adapter_tokens=64,
+        use_lcm=False,
+        use_adaln=True,
+    ):
+
+        if not isinstance(pretrained_model_path_or_dict, dict):
+            if pretrained_model_path_or_dict.endswith(".safetensors"):
+                state_dict = {"image_proj": {}, "ip_adapter": {}}
+                with safe_open(pretrained_model_path_or_dict, framework="pt", device=pipe.device) as f:
+                    for key in f.keys():
+                        if key.startswith("image_proj."):
+                            state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                        elif key.startswith("ip_adapter."):
+                            state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+            else:
+                state_dict = torch.load(pretrained_model_path_or_dict, map_location=pipe.device)
+        else:
+            state_dict = pretrained_model_path_or_dict
+        keys = list(state_dict.keys())
+        if "image_proj" not in keys and "ip_adapter" not in keys:
+            state_dict = revise_state_dict(state_dict)
+
+        # load CLIP image encoder here if it has not been registered to the pipeline yet
+        if image_encoder_or_path is not None:
+            if isinstance(image_encoder_or_path, str):
+                feature_extractor_or_path = image_encoder_or_path if feature_extractor_or_path is None else feature_extractor_or_path
+
+                image_encoder_or_path = (
+                    CLIPVisionModelWithProjection.from_pretrained(
+                        image_encoder_or_path
+                    ) if use_clip_encoder else
+                    AutoModel.from_pretrained(image_encoder_or_path)
+                )
+
+        if feature_extractor_or_path is not None:
+            if isinstance(feature_extractor_or_path, str):
+                feature_extractor_or_path = (
+                    CLIPImageProcessor() if use_clip_encoder else
+                    AutoImageProcessor.from_pretrained(feature_extractor_or_path)
+                )
+
+        # create image encoder if it has not been registered to the pipeline yet
+        if hasattr(pipe, "image_encoder") and getattr(pipe, "image_encoder", None) is None:
+            image_encoder = image_encoder_or_path.to(pipe.device, dtype=pipe.dtype)
+            pipe.register_modules(image_encoder=image_encoder)
+        else:
+            image_encoder = pipe.image_encoder
+
+        # create feature extractor if it has not been registered to the pipeline yet
+        if hasattr(pipe, "feature_extractor") and getattr(pipe, "feature_extractor", None) is None:
+            feature_extractor = feature_extractor_or_path
+            pipe.register_modules(feature_extractor=feature_extractor)
+        else:
+            feature_extractor = pipe.feature_extractor
+
+        # load adapter into unet
+        unet = getattr(pipe, pipe.unet_name) if not hasattr(pipe, "unet") else pipe.unet
+        attn_procs = init_attn_proc(unet, adapter_tokens, use_lcm, use_adaln)
+        unet.set_attn_processor(attn_procs)
+        image_proj_model = Resampler(
+            embedding_dim=image_encoder.config.hidden_size,
+            output_dim=unet.config.cross_attention_dim,
+            num_queries=adapter_tokens,
+        )
+
+        # Load pretrinaed model if needed.
+        if "ip_adapter" in state_dict.keys():
+            adapter_modules = torch.nn.ModuleList(unet.attn_processors.values())
+            missing, unexpected = adapter_modules.load_state_dict(state_dict["ip_adapter"], strict=False)
+            for mk in missing:
+                if "ln" not in mk:
+                    raise ValueError(f"Missing keys in adapter_modules: {missing}")
+        if "image_proj" in state_dict.keys():
+            image_proj_model.load_state_dict(state_dict["image_proj"])
+
+        # convert IP-Adapter Image Projection layers to diffusers
+        image_projection_layers = []
+        image_projection_layers.append(image_proj_model)
+        unet.encoder_hid_proj = MultiIPAdapterImageProjection(image_projection_layers)
+
+        # Adjust unet config to handle addtional ip hidden states.
+        unet.config.encoder_hid_dim_type = "ip_image_proj"
+        unet.to(dtype=pipe.dtype, device=pipe.device)
+
+
+def revise_state_dict(old_state_dict_or_path, map_location="cpu"):
+    new_state_dict = OrderedDict()
+    new_state_dict["image_proj"] = OrderedDict()
+    new_state_dict["ip_adapter"] = OrderedDict()
+    if isinstance(old_state_dict_or_path, str):
+        old_state_dict = torch.load(old_state_dict_or_path, map_location=map_location)
+    else:
+        old_state_dict = old_state_dict_or_path
+    for name, weight in old_state_dict.items():
+        if name.startswith("image_proj_model."):
+            new_state_dict["image_proj"][name[len("image_proj_model."):]] = weight
+        elif name.startswith("adapter_modules."):
+            new_state_dict["ip_adapter"][name[len("adapter_modules."):]] = weight
+    return new_state_dict
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+def encode_image(image_encoder, feature_extractor, image, device, num_images_per_prompt, output_hidden_states=None):
+    dtype = next(image_encoder.parameters()).dtype
+
+    if not isinstance(image, torch.Tensor):
+        image = feature_extractor(image, return_tensors="pt").pixel_values
+
+    image = image.to(device=device, dtype=dtype)
+    if output_hidden_states:
+        image_enc_hidden_states = image_encoder(image, output_hidden_states=True).hidden_states[-2]
+        image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+        return image_enc_hidden_states
+    else:
+        if isinstance(image_encoder, CLIPVisionModelWithProjection):
+            # CLIP image encoder.
+            image_embeds = image_encoder(image).image_embeds
+        else:
+            # DINO image encoder.
+            image_embeds = image_encoder(image).last_hidden_state
+        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+        return image_embeds
+
+
+def prepare_training_image_embeds(
+    image_encoder, feature_extractor,
+    ip_adapter_image, ip_adapter_image_embeds,
+    device, drop_rate, output_hidden_state, idx_to_replace=None
+):
+    if ip_adapter_image_embeds is None:
+        if not isinstance(ip_adapter_image, list):
+            ip_adapter_image = [ip_adapter_image]
+
+        # if len(ip_adapter_image) != len(unet.encoder_hid_proj.image_projection_layers):
+        #     raise ValueError(
+        #         f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+        #     )
+
+        image_embeds = []
+        for single_ip_adapter_image in ip_adapter_image:
+            if idx_to_replace is None:
+                idx_to_replace = torch.rand(len(single_ip_adapter_image)) < drop_rate
+            zero_ip_adapter_image = torch.zeros_like(single_ip_adapter_image)
+            single_ip_adapter_image[idx_to_replace] = zero_ip_adapter_image[idx_to_replace]
+            single_image_embeds = encode_image(
+                image_encoder, feature_extractor, single_ip_adapter_image, device, 1, output_hidden_state
+            )
+            single_image_embeds = torch.stack([single_image_embeds], dim=1) # FIXME
+
+            image_embeds.append(single_image_embeds)
+    else:
+        repeat_dims = [1]
+        image_embeds = []
+        for single_image_embeds in ip_adapter_image_embeds:
+            if do_classifier_free_guidance:
+                single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                single_image_embeds = single_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                )
+                single_negative_image_embeds = single_negative_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                )
+                single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+            else:
+                single_image_embeds = single_image_embeds.repeat(
+                    num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                )
+            image_embeds.append(single_image_embeds)
+
+    return image_embeds

scripts/instantir/lcm_single_step_scheduler.py

@@ -0,0 +1,537 @@
+# Copyright 2023 Stanford University Team 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.
+
+# DISCLAIMER: This code is strongly influenced by https://github.com/pesser/pytorch_diffusion
+# and https://github.com/hojonathanho/diffusion
+
+import math
+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.utils import BaseOutput, logging
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.schedulers.scheduling_utils import SchedulerMixin
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class LCMSingleStepSchedulerOutput(BaseOutput):
+    """
+    Output class for the scheduler's `step` function output.
+
+    Args:
+        pred_original_sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` for images):
+            The predicted denoised sample `(x_{0})` based on the model output from the current timestep.
+            `pred_original_sample` can be used to preview progress or for guidance.
+    """
+
+    denoised: Optional[torch.FloatTensor] = None
+
+
+# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar
+def betas_for_alpha_bar(
+    num_diffusion_timesteps,
+    max_beta=0.999,
+    alpha_transform_type="cosine",
+):
+    """
+    Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of
+    (1-beta) over time from t = [0,1].
+
+    Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up
+    to that part of the diffusion process.
+
+
+    Args:
+        num_diffusion_timesteps (`int`): the number of betas to produce.
+        max_beta (`float`): the maximum beta to use; use values lower than 1 to
+                     prevent singularities.
+        alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar.
+                     Choose from `cosine` or `exp`
+
+    Returns:
+        betas (`np.ndarray`): the betas used by the scheduler to step the model outputs
+    """
+    if alpha_transform_type == "cosine":
+
+        def alpha_bar_fn(t):
+            return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2
+
+    elif alpha_transform_type == "exp":
+
+        def alpha_bar_fn(t):
+            return math.exp(t * -12.0)
+
+    else:
+        raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}")
+
+    betas = []
+    for i in range(num_diffusion_timesteps):
+        t1 = i / num_diffusion_timesteps
+        t2 = (i + 1) / num_diffusion_timesteps
+        betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta))
+    return torch.tensor(betas, dtype=torch.float32)
+
+
+# Copied from diffusers.schedulers.scheduling_ddim.rescale_zero_terminal_snr
+def rescale_zero_terminal_snr(betas: torch.FloatTensor) -> torch.FloatTensor:
+    """
+    Rescales betas to have zero terminal SNR Based on https://arxiv.org/pdf/2305.08891.pdf (Algorithm 1)
+
+
+    Args:
+        betas (`torch.FloatTensor`):
+            the betas that the scheduler is being initialized with.
+
+    Returns:
+        `torch.FloatTensor`: rescaled betas with zero terminal SNR
+    """
+    # Convert betas to alphas_bar_sqrt
+    alphas = 1.0 - betas
+    alphas_cumprod = torch.cumprod(alphas, dim=0)
+    alphas_bar_sqrt = alphas_cumprod.sqrt()
+
+    # Store old values.
+    alphas_bar_sqrt_0 = alphas_bar_sqrt[0].clone()
+    alphas_bar_sqrt_T = alphas_bar_sqrt[-1].clone()
+
+    # Shift so the last timestep is zero.
+    alphas_bar_sqrt -= alphas_bar_sqrt_T
+
+    # Scale so the first timestep is back to the old value.
+    alphas_bar_sqrt *= alphas_bar_sqrt_0 / (alphas_bar_sqrt_0 - alphas_bar_sqrt_T)
+
+    # Convert alphas_bar_sqrt to betas
+    alphas_bar = alphas_bar_sqrt**2  # Revert sqrt
+    alphas = alphas_bar[1:] / alphas_bar[:-1]  # Revert cumprod
+    alphas = torch.cat([alphas_bar[0:1], alphas])
+    betas = 1 - alphas
+
+    return betas
+
+
+class LCMSingleStepScheduler(SchedulerMixin, ConfigMixin):
+    """
+    `LCMSingleStepScheduler` extends the denoising procedure introduced in denoising diffusion probabilistic models (DDPMs) with
+    non-Markovian guidance.
+
+    This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. [`~ConfigMixin`] takes care of storing all config
+    attributes that are passed in the scheduler's `__init__` function, such as `num_train_timesteps`. They can be
+    accessed via `scheduler.config.num_train_timesteps`. [`SchedulerMixin`] provides general loading and saving
+    functionality via the [`SchedulerMixin.save_pretrained`] and [`~SchedulerMixin.from_pretrained`] functions.
+
+    Args:
+        num_train_timesteps (`int`, defaults to 1000):
+            The number of diffusion steps to train the model.
+        beta_start (`float`, defaults to 0.0001):
+            The starting `beta` value of inference.
+        beta_end (`float`, defaults to 0.02):
+            The final `beta` value.
+        beta_schedule (`str`, defaults to `"linear"`):
+            The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from
+            `linear`, `scaled_linear`, or `squaredcos_cap_v2`.
+        trained_betas (`np.ndarray`, *optional*):
+            Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`.
+        original_inference_steps (`int`, *optional*, defaults to 50):
+            The default number of inference steps used to generate a linearly-spaced timestep schedule, from which we
+            will ultimately take `num_inference_steps` evenly spaced timesteps to form the final timestep schedule.
+        clip_sample (`bool`, defaults to `True`):
+            Clip the predicted sample for numerical stability.
+        clip_sample_range (`float`, defaults to 1.0):
+            The maximum magnitude for sample clipping. Valid only when `clip_sample=True`.
+        set_alpha_to_one (`bool`, defaults to `True`):
+            Each diffusion step uses the alphas product value at that step and at the previous one. For the final step
+            there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`,
+            otherwise it uses the alpha value at step 0.
+        steps_offset (`int`, defaults to 0):
+            An offset added to the inference steps. You can use a combination of `offset=1` and
+            `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable
+            Diffusion.
+        prediction_type (`str`, defaults to `epsilon`, *optional*):
+            Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process),
+            `sample` (directly predicts the noisy sample`) or `v_prediction` (see section 2.4 of [Imagen
+            Video](https://imagen.research.google/video/paper.pdf) paper).
+        thresholding (`bool`, defaults to `False`):
+            Whether to use the "dynamic thresholding" method. This is unsuitable for latent-space diffusion models such
+            as Stable Diffusion.
+        dynamic_thresholding_ratio (`float`, defaults to 0.995):
+            The ratio for the dynamic thresholding method. Valid only when `thresholding=True`.
+        sample_max_value (`float`, defaults to 1.0):
+            The threshold value for dynamic thresholding. Valid only when `thresholding=True`.
+        timestep_spacing (`str`, defaults to `"leading"`):
+            The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and
+            Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information.
+        timestep_scaling (`float`, defaults to 10.0):
+            The factor the timesteps will be multiplied by when calculating the consistency model boundary conditions
+            `c_skip` and `c_out`. Increasing this will decrease the approximation error (although the approximation
+            error at the default of `10.0` is already pretty small).
+        rescale_betas_zero_snr (`bool`, defaults to `False`):
+            Whether to rescale the betas to have zero terminal SNR. This enables the model to generate very bright and
+            dark samples instead of limiting it to samples with medium brightness. Loosely related to
+            [`--offset_noise`](https://github.com/huggingface/diffusers/blob/74fd735eb073eb1d774b1ab4154a0876eb82f055/examples/dreambooth/train_dreambooth.py#L506).
+    """
+
+    order = 1
+
+    @register_to_config
+    def __init__(
+        self,
+        num_train_timesteps: int = 1000,
+        beta_start: float = 0.00085,
+        beta_end: float = 0.012,
+        beta_schedule: str = "scaled_linear",
+        trained_betas: Optional[Union[np.ndarray, List[float]]] = None,
+        original_inference_steps: int = 50,
+        clip_sample: bool = False,
+        clip_sample_range: float = 1.0,
+        set_alpha_to_one: bool = True,
+        steps_offset: int = 0,
+        prediction_type: str = "epsilon",
+        thresholding: bool = False,
+        dynamic_thresholding_ratio: float = 0.995,
+        sample_max_value: float = 1.0,
+        timestep_spacing: str = "leading",
+        timestep_scaling: float = 10.0,
+        rescale_betas_zero_snr: bool = False,
+    ):
+        if trained_betas is not None:
+            self.betas = torch.tensor(trained_betas, dtype=torch.float32)
+        elif beta_schedule == "linear":
+            self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32)
+        elif beta_schedule == "scaled_linear":
+            # this schedule is very specific to the latent diffusion model.
+            self.betas = (
+                torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2
+            )
+        elif beta_schedule == "squaredcos_cap_v2":
+            # Glide cosine schedule
+            self.betas = betas_for_alpha_bar(num_train_timesteps)
+        else:
+            raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}")
+
+        # Rescale for zero SNR
+        if rescale_betas_zero_snr:
+            self.betas = rescale_zero_terminal_snr(self.betas)
+
+        self.alphas = 1.0 - self.betas
+        self.alphas_cumprod = torch.cumprod(self.alphas, dim=0)
+
+        # At every step in ddim, we are looking into the previous alphas_cumprod
+        # For the final step, there is no previous alphas_cumprod because we are already at 0
+        # `set_alpha_to_one` decides whether we set this parameter simply to one or
+        # whether we use the final alpha of the "non-previous" one.
+        self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0]
+
+        # standard deviation of the initial noise distribution
+        self.init_noise_sigma = 1.0
+
+        # setable values
+        self.num_inference_steps = None
+        self.timesteps = torch.from_numpy(np.arange(0, num_train_timesteps)[::-1].copy().astype(np.int64))
+
+        self._step_index = None
+
+    # Copied from diffusers.schedulers.scheduling_euler_discrete.EulerDiscreteScheduler._init_step_index
+    def _init_step_index(self, timestep):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.to(self.timesteps.device)
+
+        index_candidates = (self.timesteps == timestep).nonzero()
+
+        # The sigma index that is taken for the **very** first `step`
+        # is always the second index (or the last index if there is only 1)
+        # This way we can ensure we don't accidentally skip a sigma in
+        # case we start in the middle of the denoising schedule (e.g. for image-to-image)
+        if len(index_candidates) > 1:
+            step_index = index_candidates[1]
+        else:
+            step_index = index_candidates[0]
+
+        self._step_index = step_index.item()
+
+    @property
+    def step_index(self):
+        return self._step_index
+
+    def scale_model_input(self, sample: torch.FloatTensor, timestep: Optional[int] = None) -> torch.FloatTensor:
+        """
+        Ensures interchangeability with schedulers that need to scale the denoising model input depending on the
+        current timestep.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The input sample.
+            timestep (`int`, *optional*):
+                The current timestep in the diffusion chain.
+        Returns:
+            `torch.FloatTensor`:
+                A scaled input sample.
+        """
+        return sample
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler._threshold_sample
+    def _threshold_sample(self, sample: torch.FloatTensor) -> torch.FloatTensor:
+        """
+        "Dynamic thresholding: At each sampling step we set s to a certain percentile absolute pixel value in xt0 (the
+        prediction of x_0 at timestep t), and if s > 1, then we threshold xt0 to the range [-s, s] and then divide by
+        s. Dynamic thresholding pushes saturated pixels (those near -1 and 1) inwards, thereby actively preventing
+        pixels from saturation at each step. We find that dynamic thresholding results in significantly better
+        photorealism as well as better image-text alignment, especially when using very large guidance weights."
+
+        https://arxiv.org/abs/2205.11487
+        """
+        dtype = sample.dtype
+        batch_size, channels, *remaining_dims = sample.shape
+
+        if dtype not in (torch.float32, torch.float64):
+            sample = sample.float()  # upcast for quantile calculation, and clamp not implemented for cpu half
+
+        # Flatten sample for doing quantile calculation along each image
+        sample = sample.reshape(batch_size, channels * np.prod(remaining_dims))
+
+        abs_sample = sample.abs()  # "a certain percentile absolute pixel value"
+
+        s = torch.quantile(abs_sample, self.config.dynamic_thresholding_ratio, dim=1)
+        s = torch.clamp(
+            s, min=1, max=self.config.sample_max_value
+        )  # When clamped to min=1, equivalent to standard clipping to [-1, 1]
+        s = s.unsqueeze(1)  # (batch_size, 1) because clamp will broadcast along dim=0
+        sample = torch.clamp(sample, -s, s) / s  # "we threshold xt0 to the range [-s, s] and then divide by s"
+
+        sample = sample.reshape(batch_size, channels, *remaining_dims)
+        sample = sample.to(dtype)
+
+        return sample
+
+    def set_timesteps(
+        self,
+        num_inference_steps: int = None,
+        device: Union[str, torch.device] = None,
+        original_inference_steps: Optional[int] = None,
+        strength: int = 1.0,
+        timesteps: Optional[list] = None,
+    ):
+        """
+        Sets the discrete timesteps used for the diffusion chain (to be run before inference).
+
+        Args:
+            num_inference_steps (`int`):
+                The number of diffusion steps used when generating samples with a pre-trained model.
+            device (`str` or `torch.device`, *optional*):
+                The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+            original_inference_steps (`int`, *optional*):
+                The original number of inference steps, which will be used to generate a linearly-spaced timestep
+                schedule (which is different from the standard `diffusers` implementation). We will then take
+                `num_inference_steps` timesteps from this schedule, evenly spaced in terms of indices, and use that as
+                our final timestep schedule. If not set, this will default to the `original_inference_steps` attribute.
+        """
+
+        if num_inference_steps is not None and timesteps is not None:
+            raise ValueError("Can only pass one of `num_inference_steps` or `custom_timesteps`.")
+
+        if timesteps is not None:
+            for i in range(1, len(timesteps)):
+                if timesteps[i] >= timesteps[i - 1]:
+                    raise ValueError("`custom_timesteps` must be in descending order.")
+
+            if timesteps[0] >= self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`timesteps` must start before `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps}."
+                )
+
+            timesteps = np.array(timesteps, dtype=np.int64)
+        else:
+            if num_inference_steps > self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                    f" maximal {self.config.num_train_timesteps} timesteps."
+                )
+
+            self.num_inference_steps = num_inference_steps
+            original_steps = (
+                original_inference_steps if original_inference_steps is not None else self.config.original_inference_steps
+            )
+
+            if original_steps > self.config.num_train_timesteps:
+                raise ValueError(
+                    f"`original_steps`: {original_steps} cannot be larger than `self.config.train_timesteps`:"
+                    f" {self.config.num_train_timesteps} as the unet model trained with this scheduler can only handle"
+                    f" maximal {self.config.num_train_timesteps} timesteps."
+                )
+
+            if num_inference_steps > original_steps:
+                raise ValueError(
+                    f"`num_inference_steps`: {num_inference_steps} cannot be larger than `original_inference_steps`:"
+                    f" {original_steps} because the final timestep schedule will be a subset of the"
+                    f" `original_inference_steps`-sized initial timestep schedule."
+                )
+
+            # LCM Timesteps Setting
+            # Currently, only linear spacing is supported.
+            c = self.config.num_train_timesteps // original_steps
+            # LCM Training Steps Schedule
+            lcm_origin_timesteps = np.asarray(list(range(1, int(original_steps * strength) + 1))) * c - 1
+            skipping_step = len(lcm_origin_timesteps) // num_inference_steps
+            # LCM Inference Steps Schedule
+            timesteps = lcm_origin_timesteps[::-skipping_step][:num_inference_steps]
+
+        self.timesteps = torch.from_numpy(timesteps.copy()).to(device=device, dtype=torch.long)
+
+        self._step_index = None
+
+    def get_scalings_for_boundary_condition_discrete(self, timestep):
+        self.sigma_data = 0.5  # Default: 0.5
+        scaled_timestep = timestep * self.config.timestep_scaling
+
+        c_skip = self.sigma_data**2 / (scaled_timestep**2 + self.sigma_data**2)
+        c_out = scaled_timestep / (scaled_timestep**2 + self.sigma_data**2) ** 0.5
+        return c_skip, c_out
+
+    def append_dims(self, x, target_dims):
+        """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
+        dims_to_append = target_dims - x.ndim
+        if dims_to_append < 0:
+            raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
+        return x[(...,) + (None,) * dims_to_append]
+
+    def extract_into_tensor(self, a, t, x_shape):
+        b, *_ = t.shape
+        out = a.gather(-1, t)
+        return out.reshape(b, *((1,) * (len(x_shape) - 1)))
+
+    def step(
+        self,
+        model_output: torch.FloatTensor,
+        timestep: torch.Tensor,
+        sample: torch.FloatTensor,
+        generator: Optional[torch.Generator] = None,
+        return_dict: bool = True,
+    ) -> Union[LCMSingleStepSchedulerOutput, Tuple]:
+        """
+        Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion
+        process from the learned model outputs (most often the predicted noise).
+
+        Args:
+            model_output (`torch.FloatTensor`):
+                The direct output from learned diffusion model.
+            timestep (`float`):
+                The current discrete timestep in the diffusion chain.
+            sample (`torch.FloatTensor`):
+                A current instance of a sample created by the diffusion process.
+            generator (`torch.Generator`, *optional*):
+                A random number generator.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] or `tuple`.
+        Returns:
+            [`~schedulers.scheduling_utils.LCMSchedulerOutput`] or `tuple`:
+                If return_dict is `True`, [`~schedulers.scheduling_lcm.LCMSchedulerOutput`] is returned, otherwise a
+                tuple is returned where the first element is the sample tensor.
+        """
+        # 0. make sure everything is on the same device
+        alphas_cumprod = self.alphas_cumprod.to(sample.device)
+
+        # 1. compute alphas, betas
+        if timestep.ndim == 0:
+            timestep = timestep.unsqueeze(0)
+        alpha_prod_t = self.extract_into_tensor(alphas_cumprod, timestep, sample.shape)
+        beta_prod_t = 1 - alpha_prod_t
+
+        # 2. Get scalings for boundary conditions
+        c_skip, c_out = self.get_scalings_for_boundary_condition_discrete(timestep)
+        c_skip, c_out = [self.append_dims(x, sample.ndim) for x in [c_skip, c_out]]
+
+        # 3. Compute the predicted original sample x_0 based on the model parameterization
+        if self.config.prediction_type == "epsilon":  # noise-prediction
+            predicted_original_sample = (sample - torch.sqrt(beta_prod_t) * model_output) / torch.sqrt(alpha_prod_t)
+        elif self.config.prediction_type == "sample":  # x-prediction
+            predicted_original_sample = model_output
+        elif self.config.prediction_type == "v_prediction":  # v-prediction
+            predicted_original_sample = torch.sqrt(alpha_prod_t) * sample - torch.sqrt(beta_prod_t) * model_output
+        else:
+            raise ValueError(
+                f"prediction_type given as {self.config.prediction_type} must be one of `epsilon`, `sample` or"
+                " `v_prediction` for `LCMScheduler`."
+            )
+
+        # 4. Clip or threshold "predicted x_0"
+        if self.config.thresholding:
+            predicted_original_sample = self._threshold_sample(predicted_original_sample)
+        elif self.config.clip_sample:
+            predicted_original_sample = predicted_original_sample.clamp(
+                -self.config.clip_sample_range, self.config.clip_sample_range
+            )
+
+        # 5. Denoise model output using boundary conditions
+        denoised = c_out * predicted_original_sample + c_skip * sample
+
+        if not return_dict:
+            return (denoised, )
+
+        return LCMSingleStepSchedulerOutput(denoised=denoised)
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise
+    def add_noise(
+        self,
+        original_samples: torch.FloatTensor,
+        noise: torch.FloatTensor,
+        timesteps: torch.IntTensor,
+    ) -> torch.FloatTensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as original_samples
+        alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device, dtype=original_samples.dtype)
+        timesteps = timesteps.to(original_samples.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
+        return noisy_samples
+
+    # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.get_velocity
+    def get_velocity(
+        self, sample: torch.FloatTensor, noise: torch.FloatTensor, timesteps: torch.IntTensor
+    ) -> torch.FloatTensor:
+        # Make sure alphas_cumprod and timestep have same device and dtype as sample
+        alphas_cumprod = self.alphas_cumprod.to(device=sample.device, dtype=sample.dtype)
+        timesteps = timesteps.to(sample.device)
+
+        sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5
+        sqrt_alpha_prod = sqrt_alpha_prod.flatten()
+        while len(sqrt_alpha_prod.shape) < len(sample.shape):
+            sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
+
+        sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5
+        sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
+        while len(sqrt_one_minus_alpha_prod.shape) < len(sample.shape):
+            sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
+
+        velocity = sqrt_alpha_prod * noise - sqrt_one_minus_alpha_prod * sample
+        return velocity
+
+    def __len__(self):
+        return self.config.num_train_timesteps