animatediff

__assets__/docs/animatediff.md

+# AnimateDiff: training and inference setup
+## Setups for Inference
+
+### Prepare Environment
+
+***We updated our inference code with xformers and a sequential decoding trick. Now AnimateDiff takes only ~12GB VRAM to inference, and run on a single RTX3090 !!***
+
+```
+git clone https://github.com/guoyww/AnimateDiff.git
+cd AnimateDiff
+
+conda env create -f environment.yaml
+conda activate animatediff
+```
+
+### Download Base T2I & Motion Module Checkpoints
+We provide two versions of our Motion Module, which are trained on stable-diffusion-v1-4 and finetuned on v1-5 seperately.
+It's recommanded to try both of them for best results.
+```
+git lfs install
+git clone https://huggingface.co/runwayml/stable-diffusion-v1-5 models/StableDiffusion/
+
+bash download_bashscripts/0-MotionModule.sh
+```
+You may also directly download the motion module checkpoints from [Google Drive](https://drive.google.com/drive/folders/1EqLC65eR1-W-sGD0Im7fkED6c8GkiNFI?usp=sharing) / [HuggingFace](https://huggingface.co/guoyww/animatediff) / [CivitAI](https://civitai.com/models/108836/animatediff-motion-modules), then put them in `models/Motion_Module/` folder.
+
+### Prepare Personalize T2I
+Here we provide inference configs for 6 demo T2I on CivitAI.
+You may run the following bash scripts to download these checkpoints.
+```
+bash download_bashscripts/1-ToonYou.sh
+bash download_bashscripts/2-Lyriel.sh
+bash download_bashscripts/3-RcnzCartoon.sh
+bash download_bashscripts/4-MajicMix.sh
+bash download_bashscripts/5-RealisticVision.sh
+bash download_bashscripts/6-Tusun.sh
+bash download_bashscripts/7-FilmVelvia.sh
+bash download_bashscripts/8-GhibliBackground.sh
+```
+
+### Inference
+After downloading the above peronalized T2I checkpoints, run the following commands to generate animations. The results will automatically be saved to `samples/` folder.
+```
+python -m scripts.animate --config configs/prompts/1-ToonYou.yaml
+python -m scripts.animate --config configs/prompts/2-Lyriel.yaml
+python -m scripts.animate --config configs/prompts/3-RcnzCartoon.yaml
+python -m scripts.animate --config configs/prompts/4-MajicMix.yaml
+python -m scripts.animate --config configs/prompts/5-RealisticVision.yaml
+python -m scripts.animate --config configs/prompts/6-Tusun.yaml
+python -m scripts.animate --config configs/prompts/7-FilmVelvia.yaml
+python -m scripts.animate --config configs/prompts/8-GhibliBackground.yaml
+```
+
+To generate animations with a new DreamBooth/LoRA model, you may create a new config `.yaml` file in the following format:
+```
+- inference_config: "[path to motion module config file]"
+
+  motion_module:
+    - "models/Motion_Module/mm_sd_v14.ckpt"
+    - "models/Motion_Module/mm_sd_v15.ckpt"
+    
+    motion_module_lora_configs:
+    - path:  "[path to MotionLoRA model]"
+      alpha: 1.0
+    - ...
+
+  dreambooth_path: "[path to your DreamBooth model .safetensors file]"
+  lora_model_path: "[path to your LoRA model .safetensors file, leave it empty string if not needed]"
+
+  steps:          25
+  guidance_scale: 7.5
+
+  prompt:
+    - "[positive prompt]"
+
+  n_prompt:
+    - "[negative prompt]"
+```
+Then run the following commands:
+```
+python -m scripts.animate --config [path to the config file]
+```
+
+
+## Steps for Training
+
+### Dataset
+Before training, download the videos files and the `.csv` annotations of [WebVid10M](https://maxbain.com/webvid-dataset/) to the local mechine.
+Note that our examplar training script requires all the videos to be saved in a single folder. You may change this by modifying `animatediff/data/dataset.py`.
+
+### Configuration
+After dataset preparations, update the below data paths in the config `.yaml` files in `configs/training/` folder:
+```
+train_data:
+  csv_path:     [Replace with .csv Annotation File Path]
+  video_folder: [Replace with Video Folder Path]
+  sample_size:  256
+```
+Other training parameters (lr, epochs, validation settings, etc.) are also included in the config files.
+
+### Training
+To finetune the unet's image layers
+```
+torchrun --nnodes=1 --nproc_per_node=1 train.py --config configs/training/v1/image_finetune.yaml
+```
+
+To train motion modules
+```
+torchrun --nnodes=1 --nproc_per_node=1 train.py --config configs/training/v1/training.yaml
+```

__assets__/docs/gallery.md

+# Gallery
+Here we demonstrate several best results we found in our experiments.
+
+<table class="center">
+    <tr>
+    <td><img src="../animations/model_01/01.gif"></td>
+    <td><img src="../animations/model_01/02.gif"></td>
+    <td><img src="../animations/model_01/03.gif"></td>
+    <td><img src="../animations/model_01/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/30240/toonyou">ToonYou</a></p>
+
+<table>
+    <tr>
+    <td><img src="../animations/model_02/01.gif"></td>
+    <td><img src="../animations/model_02/02.gif"></td>
+    <td><img src="../animations/model_02/03.gif"></td>
+    <td><img src="../animations/model_02/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/4468/counterfeit-v30">Counterfeit V3.0</a></p>
+
+<table>
+    <tr>
+    <td><img src="../animations/model_03/01.gif"></td>
+    <td><img src="../animations/model_03/02.gif"></td>
+    <td><img src="../animations/model_03/03.gif"></td>
+    <td><img src="../animations/model_03/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/4201/realistic-vision-v20">Realistic Vision V2.0</a></p>
+
+<table>
+    <tr>
+    <td><img src="../animations/model_04/01.gif"></td>
+    <td><img src="../animations/model_04/02.gif"></td>
+    <td><img src="../animations/model_04/03.gif"></td>
+    <td><img src="../animations/model_04/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model： <a href="https://civitai.com/models/43331/majicmix-realistic">majicMIX Realistic</a></p>
+
+<table>
+    <tr>
+    <td><img src="../animations/model_05/01.gif"></td>
+    <td><img src="../animations/model_05/02.gif"></td>
+    <td><img src="../animations/model_05/03.gif"></td>
+    <td><img src="../animations/model_05/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/66347/rcnz-cartoon-3d">RCNZ Cartoon</a></p>
+
+<table>
+    <tr>
+    <td><img src="../animations/model_06/01.gif"></td>
+    <td><img src="../animations/model_06/02.gif"></td>
+    <td><img src="../animations/model_06/03.gif"></td>
+    <td><img src="../animations/model_06/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">Model：<a href="https://civitai.com/models/33208/filmgirl-film-grain-lora-and-loha">FilmVelvia</a></p>
+
+#### Community Cases
+Here are some samples contributed by the community artists. Create a Pull Request if you would like to show your results here😚.
+
+<table>
+    <tr>
+    <td><img src="../animations/model_07/init.jpg"></td>
+    <td><img src="../animations/model_07/01.gif"></td>
+    <td><img src="../animations/model_07/02.gif"></td>
+    <td><img src="../animations/model_07/03.gif"></td>
+    <td><img src="../animations/model_07/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">
+Character Model：<a href="https://civitai.com/models/13237/genshen-impact-yoimiya">Yoimiya</a> 
+(with an initial reference image, see <a href="https://github.com/talesofai/AnimateDiff">WIP fork</a> for the extended implementation.)
+
+
+<table>
+    <tr>
+    <td><img src="../animations/model_08/01.gif"></td>
+    <td><img src="../animations/model_08/02.gif"></td>
+    <td><img src="../animations/model_08/03.gif"></td>
+    <td><img src="../animations/model_08/04.gif"></td>
+    </tr>
+</table>
+<p style="margin-left: 2em; margin-top: -1em">
+Character Model：<a href="https://civitai.com/models/9850/paimon-genshin-impact">Paimon</a>;
+Pose Model：<a href="https://civitai.com/models/107295/or-holdingsign">Hold Sign</a></p>
+
+

animatediff/data/dataset.py

+import os, io, csv, math, random
+import numpy as np
+from einops import rearrange
+from decord import VideoReader
+
+import torch
+import torchvision.transforms as transforms
+from torch.utils.data.dataset import Dataset
+from animatediff.utils.util import zero_rank_print
+
+
+
+class WebVid10M(Dataset):
+    def __init__(
+            self,
+            csv_path, video_folder,
+            sample_size=256, sample_stride=4, sample_n_frames=16,
+            is_image=False,
+        ):
+        zero_rank_print(f"loading annotations from {csv_path} ...")
+        with open(csv_path, 'r') as csvfile:
+            self.dataset = list(csv.DictReader(csvfile))
+        self.length = len(self.dataset)
+        zero_rank_print(f"data scale: {self.length}")
+
+        self.video_folder    = video_folder
+        self.sample_stride   = sample_stride
+        self.sample_n_frames = sample_n_frames
+        self.is_image        = is_image
+        
+        sample_size = tuple(sample_size) if not isinstance(sample_size, int) else (sample_size, sample_size)
+        self.pixel_transforms = transforms.Compose([
+            transforms.RandomHorizontalFlip(),
+            transforms.Resize(sample_size[0]),
+            transforms.CenterCrop(sample_size),
+            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5], inplace=True),
+        ])
+    
+    def get_batch(self, idx):
+        video_dict = self.dataset[idx]
+        videoid, name, page_dir = video_dict['videoid'], video_dict['name'], video_dict['page_dir']
+        
+        video_dir    = os.path.join(self.video_folder, f"{videoid}.mp4")
+        video_reader = VideoReader(video_dir)
+        video_length = len(video_reader)
+        
+        if not self.is_image:
+            clip_length = min(video_length, (self.sample_n_frames - 1) * self.sample_stride + 1)
+            start_idx   = random.randint(0, video_length - clip_length)
+            batch_index = np.linspace(start_idx, start_idx + clip_length - 1, self.sample_n_frames, dtype=int)
+        else:
+            batch_index = [random.randint(0, video_length - 1)]
+
+        pixel_values = torch.from_numpy(video_reader.get_batch(batch_index).asnumpy()).permute(0, 3, 1, 2).contiguous()
+        pixel_values = pixel_values / 255.
+        del video_reader
+
+        if self.is_image:
+            pixel_values = pixel_values[0]
+        
+        return pixel_values, name
+
+    def __len__(self):
+        return self.length
+
+    def __getitem__(self, idx):
+        while True:
+            try:
+                pixel_values, name = self.get_batch(idx)
+                break
+
+            except Exception as e:
+                idx = random.randint(0, self.length-1)
+
+        pixel_values = self.pixel_transforms(pixel_values)
+        sample = dict(pixel_values=pixel_values, text=name)
+        return sample
+
+
+
+if __name__ == "__main__":
+    from animatediff.utils.util import save_videos_grid
+
+    dataset = WebVid10M(
+        csv_path="/mnt/petrelfs/guoyuwei/projects/datasets/webvid/results_2M_val.csv",
+        video_folder="/mnt/petrelfs/guoyuwei/projects/datasets/webvid/2M_val",
+        sample_size=256,
+        sample_stride=4, sample_n_frames=16,
+        is_image=True,
+    )
+    import pdb
+    pdb.set_trace()
+    
+    dataloader = torch.utils.data.DataLoader(dataset, batch_size=4, num_workers=16,)
+    for idx, batch in enumerate(dataloader):
+        print(batch["pixel_values"].shape, len(batch["text"]))
+        # for i in range(batch["pixel_values"].shape[0]):
+        #     save_videos_grid(batch["pixel_values"][i:i+1].permute(0,2,1,3,4), os.path.join(".", f"{idx}-{i}.mp4"), rescale=True)

animatediff/models/attention.py

+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from dataclasses import dataclass
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.models.attention import CrossAttention, FeedForward, AdaLayerNorm
+
+from einops import rearrange, repeat
+import pdb
+
+@dataclass
+class Transformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+class Transformer3DModel(ModelMixin, ConfigMixin):
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+
+        unet_use_cross_frame_attention=None,
+        unet_use_temporal_attention=None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        # Define input layers
+        self.in_channels = in_channels
+
+        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        if use_linear_projection:
+            self.proj_in = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_in = nn.Conv2d(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+
+        # Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    upcast_attention=upcast_attention,
+
+                    unet_use_cross_frame_attention=unet_use_cross_frame_attention,
+                    unet_use_temporal_attention=unet_use_temporal_attention,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        if use_linear_projection:
+            self.proj_out = nn.Linear(in_channels, inner_dim)
+        else:
+            self.proj_out = nn.Conv2d(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, return_dict: bool = True):
+        # Input
+        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+        encoder_hidden_states = repeat(encoder_hidden_states, 'b n c -> (b f) n c', f=video_length)
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+            hidden_states = self.proj_in(hidden_states)
+
+        # Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                timestep=timestep,
+                video_length=video_length
+            )
+
+        # Output
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+
+        output = hidden_states + residual
+
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        if not return_dict:
+            return (output,)
+
+        return Transformer3DModelOutput(sample=output)
+
+
+class BasicTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        upcast_attention: bool = False,
+
+        unet_use_cross_frame_attention = None,
+        unet_use_temporal_attention = None,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+        self.use_ada_layer_norm = num_embeds_ada_norm is not None
+        self.unet_use_cross_frame_attention = unet_use_cross_frame_attention
+        self.unet_use_temporal_attention = unet_use_temporal_attention
+
+        # SC-Attn
+        assert unet_use_cross_frame_attention is not None
+        if unet_use_cross_frame_attention:
+            self.attn1 = SparseCausalAttention2D(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn1 = CrossAttention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+        # Cross-Attn
+        if cross_attention_dim is not None:
+            self.attn2 = CrossAttention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+        else:
+            self.attn2 = None
+
+        if cross_attention_dim is not None:
+            self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+        else:
+            self.norm2 = None
+
+        # Feed-forward
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.norm3 = nn.LayerNorm(dim)
+
+        # Temp-Attn
+        assert unet_use_temporal_attention is not None
+        if unet_use_temporal_attention:
+            self.attn_temp = CrossAttention(
+                query_dim=dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+            )
+            nn.init.zeros_(self.attn_temp.to_out[0].weight.data)
+            self.norm_temp = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+    def set_use_memory_efficient_attention_xformers(self, use_memory_efficient_attention_xformers: bool):
+        if not is_xformers_available():
+            print("Here is how to install it")
+            raise ModuleNotFoundError(
+                "Refer to https://github.com/facebookresearch/xformers for more information on how to install"
+                " xformers",
+                name="xformers",
+            )
+        elif not torch.cuda.is_available():
+            raise ValueError(
+                "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is only"
+                " available for GPU "
+            )
+        else:
+            try:
+                # Make sure we can run the memory efficient attention
+                _ = xformers.ops.memory_efficient_attention(
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                    torch.randn((1, 2, 40), device="cuda"),
+                )
+            except Exception as e:
+                raise e
+            self.attn1._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            if self.attn2 is not None:
+                self.attn2._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+            # self.attn_temp._use_memory_efficient_attention_xformers = use_memory_efficient_attention_xformers
+
+    def forward(self, hidden_states, encoder_hidden_states=None, timestep=None, attention_mask=None, video_length=None):
+        # SparseCausal-Attention
+        norm_hidden_states = (
+            self.norm1(hidden_states, timestep) if self.use_ada_layer_norm else self.norm1(hidden_states)
+        )
+
+        # if self.only_cross_attention:
+        #     hidden_states = (
+        #         self.attn1(norm_hidden_states, encoder_hidden_states, attention_mask=attention_mask) + hidden_states
+        #     )
+        # else:
+        #     hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states
+
+        # pdb.set_trace()
+        if self.unet_use_cross_frame_attention:
+            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask, video_length=video_length) + hidden_states
+        else:
+            hidden_states = self.attn1(norm_hidden_states, attention_mask=attention_mask) + hidden_states
+
+        if self.attn2 is not None:
+            # Cross-Attention
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(hidden_states)
+            )
+            hidden_states = (
+                self.attn2(
+                    norm_hidden_states, encoder_hidden_states=encoder_hidden_states, attention_mask=attention_mask
+                )
+                + hidden_states
+            )
+
+        # Feed-forward
+        hidden_states = self.ff(self.norm3(hidden_states)) + hidden_states
+
+        # Temporal-Attention
+        if self.unet_use_temporal_attention:
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length)
+            norm_hidden_states = (
+                self.norm_temp(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_temp(hidden_states)
+            )
+            hidden_states = self.attn_temp(norm_hidden_states) + hidden_states
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

animatediff/models/motion_module.py

+from dataclasses import dataclass
+from typing import List, Optional, Tuple, Union
+
+import torch
+import numpy as np
+import torch.nn.functional as F
+from torch import nn
+import torchvision
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.modeling_utils import ModelMixin
+from diffusers.utils import BaseOutput
+from diffusers.utils.import_utils import is_xformers_available
+from diffusers.models.attention import CrossAttention, FeedForward
+
+from einops import rearrange, repeat
+import math
+
+
+def zero_module(module):
+    # Zero out the parameters of a module and return it.
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+@dataclass
+class TemporalTransformer3DModelOutput(BaseOutput):
+    sample: torch.FloatTensor
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+def get_motion_module(
+    in_channels,
+    motion_module_type: str, 
+    motion_module_kwargs: dict
+):
+    if motion_module_type == "Vanilla":
+        return VanillaTemporalModule(in_channels=in_channels, **motion_module_kwargs,)    
+    else:
+        raise ValueError
+
+
+class VanillaTemporalModule(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads                = 8,
+        num_transformer_block              = 2,
+        attention_block_types              =( "Temporal_Self", "Temporal_Self" ),
+        cross_frame_attention_mode         = None,
+        temporal_position_encoding         = False,
+        temporal_position_encoding_max_len = 24,
+        temporal_attention_dim_div         = 1,
+        zero_initialize                    = True,
+    ):
+        super().__init__()
+        
+        self.temporal_transformer = TemporalTransformer3DModel(
+            in_channels=in_channels,
+            num_attention_heads=num_attention_heads,
+            attention_head_dim=in_channels // num_attention_heads // temporal_attention_dim_div,
+            num_layers=num_transformer_block,
+            attention_block_types=attention_block_types,
+            cross_frame_attention_mode=cross_frame_attention_mode,
+            temporal_position_encoding=temporal_position_encoding,
+            temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+        )
+        
+        if zero_initialize:
+            self.temporal_transformer.proj_out = zero_module(self.temporal_transformer.proj_out)
+
+    def forward(self, input_tensor, temb, encoder_hidden_states, attention_mask=None, anchor_frame_idx=None):
+        hidden_states = input_tensor
+        hidden_states = self.temporal_transformer(hidden_states, encoder_hidden_states, attention_mask)
+
+        output = hidden_states
+        return output
+
+
+class TemporalTransformer3DModel(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        num_attention_heads,
+        attention_head_dim,
+
+        num_layers,
+        attention_block_types              = ( "Temporal_Self", "Temporal_Self", ),        
+        dropout                            = 0.0,
+        norm_num_groups                    = 32,
+        cross_attention_dim                = 768,
+        activation_fn                      = "geglu",
+        attention_bias                     = False,
+        upcast_attention                   = False,
+        
+        cross_frame_attention_mode         = None,
+        temporal_position_encoding         = False,
+        temporal_position_encoding_max_len = 24,
+    ):
+        super().__init__()
+
+        inner_dim = num_attention_heads * attention_head_dim
+
+        self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+        self.proj_in = nn.Linear(in_channels, inner_dim)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                TemporalTransformerBlock(
+                    dim=inner_dim,
+                    num_attention_heads=num_attention_heads,
+                    attention_head_dim=attention_head_dim,
+                    attention_block_types=attention_block_types,
+                    dropout=dropout,
+                    norm_num_groups=norm_num_groups,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    attention_bias=attention_bias,
+                    upcast_attention=upcast_attention,
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+                for d in range(num_layers)
+            ]
+        )
+        self.proj_out = nn.Linear(inner_dim, in_channels)    
+    
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None):
+        assert hidden_states.dim() == 5, f"Expected hidden_states to have ndim=5, but got ndim={hidden_states.dim()}."
+        video_length = hidden_states.shape[2]
+        hidden_states = rearrange(hidden_states, "b c f h w -> (b f) c h w")
+
+        batch, channel, height, weight = hidden_states.shape
+        residual = hidden_states
+
+        hidden_states = self.norm(hidden_states)
+        inner_dim = hidden_states.shape[1]
+        hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * weight, inner_dim)
+        hidden_states = self.proj_in(hidden_states)
+
+        # Transformer Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(hidden_states, encoder_hidden_states=encoder_hidden_states, video_length=video_length)
+        
+        # output
+        hidden_states = self.proj_out(hidden_states)
+        hidden_states = hidden_states.reshape(batch, height, weight, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+        output = hidden_states + residual
+        output = rearrange(output, "(b f) c h w -> b c f h w", f=video_length)
+        
+        return output
+
+
+class TemporalTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_attention_heads,
+        attention_head_dim,
+        attention_block_types              = ( "Temporal_Self", "Temporal_Self", ),
+        dropout                            = 0.0,
+        norm_num_groups                    = 32,
+        cross_attention_dim                = 768,
+        activation_fn                      = "geglu",
+        attention_bias                     = False,
+        upcast_attention                   = False,
+        cross_frame_attention_mode         = None,
+        temporal_position_encoding         = False,
+        temporal_position_encoding_max_len = 24,
+    ):
+        super().__init__()
+
+        attention_blocks = []
+        norms = []
+        
+        for block_name in attention_block_types:
+            attention_blocks.append(
+                VersatileAttention(
+                    attention_mode=block_name.split("_")[0],
+                    cross_attention_dim=cross_attention_dim if block_name.endswith("_Cross") else None,
+                    
+                    query_dim=dim,
+                    heads=num_attention_heads,
+                    dim_head=attention_head_dim,
+                    dropout=dropout,
+                    bias=attention_bias,
+                    upcast_attention=upcast_attention,
+        
+                    cross_frame_attention_mode=cross_frame_attention_mode,
+                    temporal_position_encoding=temporal_position_encoding,
+                    temporal_position_encoding_max_len=temporal_position_encoding_max_len,
+                )
+            )
+            norms.append(nn.LayerNorm(dim))
+            
+        self.attention_blocks = nn.ModuleList(attention_blocks)
+        self.norms = nn.ModuleList(norms)
+
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn)
+        self.ff_norm = nn.LayerNorm(dim)
+
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None):
+        for attention_block, norm in zip(self.attention_blocks, self.norms):
+            norm_hidden_states = norm(hidden_states)
+            hidden_states = attention_block(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states if attention_block.is_cross_attention else None,
+                video_length=video_length,
+            ) + hidden_states
+            
+        hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
+        
+        output = hidden_states  
+        return output
+
+
+class PositionalEncoding(nn.Module):
+    def __init__(
+        self, 
+        d_model, 
+        dropout = 0., 
+        max_len = 24
+    ):
+        super().__init__()
+        self.dropout = nn.Dropout(p=dropout)
+        position = torch.arange(max_len).unsqueeze(1)
+        div_term = torch.exp(torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model))
+        pe = torch.zeros(1, max_len, d_model)
+        pe[0, :, 0::2] = torch.sin(position * div_term)
+        pe[0, :, 1::2] = torch.cos(position * div_term)
+        self.register_buffer('pe', pe)
+
+    def forward(self, x):
+        x = x + self.pe[:, :x.size(1)]
+        return self.dropout(x)
+
+
+class VersatileAttention(CrossAttention):
+    def __init__(
+            self,
+            attention_mode                     = None,
+            cross_frame_attention_mode         = None,
+            temporal_position_encoding         = False,
+            temporal_position_encoding_max_len = 24,            
+            *args, **kwargs
+        ):
+        super().__init__(*args, **kwargs)
+        assert attention_mode == "Temporal"
+
+        self.attention_mode = attention_mode
+        self.is_cross_attention = kwargs["cross_attention_dim"] is not None
+        
+        self.pos_encoder = PositionalEncoding(
+            kwargs["query_dim"],
+            dropout=0., 
+            max_len=temporal_position_encoding_max_len
+        ) if (temporal_position_encoding and attention_mode == "Temporal") else None
+
+    def extra_repr(self):
+        return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, video_length=None):
+        batch_size, sequence_length, _ = hidden_states.shape
+
+        if self.attention_mode == "Temporal":
+            d = hidden_states.shape[1]
+            hidden_states = rearrange(hidden_states, "(b f) d c -> (b d) f c", f=video_length)
+            
+            if self.pos_encoder is not None:
+                hidden_states = self.pos_encoder(hidden_states)
+            
+            encoder_hidden_states = repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d) if encoder_hidden_states is not None else encoder_hidden_states
+        else:
+            raise NotImplementedError
+
+        encoder_hidden_states = encoder_hidden_states
+
+        if self.group_norm is not None:
+            hidden_states = self.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = self.to_q(hidden_states)
+        dim = query.shape[-1]
+        query = self.reshape_heads_to_batch_dim(query)
+
+        if self.added_kv_proj_dim is not None:
+            raise NotImplementedError
+
+        encoder_hidden_states = encoder_hidden_states if encoder_hidden_states is not None else hidden_states
+        key = self.to_k(encoder_hidden_states)
+        value = self.to_v(encoder_hidden_states)
+
+        key = self.reshape_heads_to_batch_dim(key)
+        value = self.reshape_heads_to_batch_dim(value)
+
+        if attention_mask is not None:
+            if attention_mask.shape[-1] != query.shape[1]:
+                target_length = query.shape[1]
+                attention_mask = F.pad(attention_mask, (0, target_length), value=0.0)
+                attention_mask = attention_mask.repeat_interleave(self.heads, dim=0)
+
+        # attention, what we cannot get enough of
+        if self._use_memory_efficient_attention_xformers:
+            hidden_states = self._memory_efficient_attention_xformers(query, key, value, attention_mask)
+            # Some versions of xformers return output in fp32, cast it back to the dtype of the input
+            hidden_states = hidden_states.to(query.dtype)
+        else:
+            if self._slice_size is None or query.shape[0] // self._slice_size == 1:
+                hidden_states = self._attention(query, key, value, attention_mask)
+            else:
+                hidden_states = self._sliced_attention(query, key, value, sequence_length, dim, attention_mask)
+
+        # linear proj
+        hidden_states = self.to_out[0](hidden_states)
+
+        # dropout
+        hidden_states = self.to_out[1](hidden_states)
+
+        if self.attention_mode == "Temporal":
+            hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
+
+        return hidden_states

animatediff/models/resnet.py

+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/resnet.py
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from einops import rearrange
+
+
+class InflatedConv3d(nn.Conv2d):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class InflatedGroupNorm(nn.GroupNorm):
+    def forward(self, x):
+        video_length = x.shape[2]
+
+        x = rearrange(x, "b c f h w -> (b f) c h w")
+        x = super().forward(x)
+        x = rearrange(x, "(b f) c h w -> b c f h w", f=video_length)
+
+        return x
+
+
+class Upsample3D(nn.Module):
+    def __init__(self, channels, use_conv=False, use_conv_transpose=False, out_channels=None, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.use_conv_transpose = use_conv_transpose
+        self.name = name
+
+        conv = None
+        if use_conv_transpose:
+            raise NotImplementedError
+        elif use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, padding=1)
+
+    def forward(self, hidden_states, output_size=None):
+        assert hidden_states.shape[1] == self.channels
+
+        if self.use_conv_transpose:
+            raise NotImplementedError
+
+        # Cast to float32 to as 'upsample_nearest2d_out_frame' op does not support bfloat16
+        dtype = hidden_states.dtype
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(torch.float32)
+
+        # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+        if hidden_states.shape[0] >= 64:
+            hidden_states = hidden_states.contiguous()
+
+        # if `output_size` is passed we force the interpolation output
+        # size and do not make use of `scale_factor=2`
+        if output_size is None:
+            hidden_states = F.interpolate(hidden_states, scale_factor=[1.0, 2.0, 2.0], mode="nearest")
+        else:
+            hidden_states = F.interpolate(hidden_states, size=output_size, mode="nearest")
+
+        # If the input is bfloat16, we cast back to bfloat16
+        if dtype == torch.bfloat16:
+            hidden_states = hidden_states.to(dtype)
+
+        # if self.use_conv:
+        #     if self.name == "conv":
+        #         hidden_states = self.conv(hidden_states)
+        #     else:
+        #         hidden_states = self.Conv2d_0(hidden_states)
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class Downsample3D(nn.Module):
+    def __init__(self, channels, use_conv=False, out_channels=None, padding=1, name="conv"):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+        self.use_conv = use_conv
+        self.padding = padding
+        stride = 2
+        self.name = name
+
+        if use_conv:
+            self.conv = InflatedConv3d(self.channels, self.out_channels, 3, stride=stride, padding=padding)
+        else:
+            raise NotImplementedError
+
+    def forward(self, hidden_states):
+        assert hidden_states.shape[1] == self.channels
+        if self.use_conv and self.padding == 0:
+            raise NotImplementedError
+
+        assert hidden_states.shape[1] == self.channels
+        hidden_states = self.conv(hidden_states)
+
+        return hidden_states
+
+
+class ResnetBlock3D(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout=0.0,
+        temb_channels=512,
+        groups=32,
+        groups_out=None,
+        pre_norm=True,
+        eps=1e-6,
+        non_linearity="swish",
+        time_embedding_norm="default",
+        output_scale_factor=1.0,
+        use_in_shortcut=None,
+        use_inflated_groupnorm=False,
+    ):
+        super().__init__()
+        self.pre_norm = pre_norm
+        self.pre_norm = True
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+        self.time_embedding_norm = time_embedding_norm
+        self.output_scale_factor = output_scale_factor
+
+        if groups_out is None:
+            groups_out = groups
+
+        assert use_inflated_groupnorm != None
+        if use_inflated_groupnorm:
+            self.norm1 = InflatedGroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+        else:
+            self.norm1 = torch.nn.GroupNorm(num_groups=groups, num_channels=in_channels, eps=eps, affine=True)
+
+        self.conv1 = InflatedConv3d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if temb_channels is not None:
+            if self.time_embedding_norm == "default":
+                time_emb_proj_out_channels = out_channels
+            elif self.time_embedding_norm == "scale_shift":
+                time_emb_proj_out_channels = out_channels * 2
+            else:
+                raise ValueError(f"unknown time_embedding_norm : {self.time_embedding_norm} ")
+
+            self.time_emb_proj = torch.nn.Linear(temb_channels, time_emb_proj_out_channels)
+        else:
+            self.time_emb_proj = None
+
+        if use_inflated_groupnorm:
+            self.norm2 = InflatedGroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+        else:
+            self.norm2 = torch.nn.GroupNorm(num_groups=groups_out, num_channels=out_channels, eps=eps, affine=True)
+
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = InflatedConv3d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+
+        if non_linearity == "swish":
+            self.nonlinearity = lambda x: F.silu(x)
+        elif non_linearity == "mish":
+            self.nonlinearity = Mish()
+        elif non_linearity == "silu":
+            self.nonlinearity = nn.SiLU()
+
+        self.use_in_shortcut = self.in_channels != self.out_channels if use_in_shortcut is None else use_in_shortcut
+
+        self.conv_shortcut = None
+        if self.use_in_shortcut:
+            self.conv_shortcut = InflatedConv3d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, input_tensor, temb):
+        hidden_states = input_tensor
+
+        hidden_states = self.norm1(hidden_states)
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.conv1(hidden_states)
+
+        if temb is not None:
+            temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None, None]
+
+        if temb is not None and self.time_embedding_norm == "default":
+            hidden_states = hidden_states + temb
+
+        hidden_states = self.norm2(hidden_states)
+
+        if temb is not None and self.time_embedding_norm == "scale_shift":
+            scale, shift = torch.chunk(temb, 2, dim=1)
+            hidden_states = hidden_states * (1 + scale) + shift
+
+        hidden_states = self.nonlinearity(hidden_states)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.conv2(hidden_states)
+
+        if self.conv_shortcut is not None:
+            input_tensor = self.conv_shortcut(input_tensor)
+
+        output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+        return output_tensor
+
+
+class Mish(torch.nn.Module):
+    def forward(self, hidden_states):
+        return hidden_states * torch.tanh(torch.nn.functional.softplus(hidden_states))
\ No newline at end of file