diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml
index 8e57220c54ad..81d7a95454d7 100644
--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@@ -267,6 +267,8 @@
title: HunyuanDiT2DControlNetModel
- local: api/models/controlnet_sd3
title: SD3ControlNetModel
+ - local: api/models/controlnet_sparsectrl
+ title: SparseControlNetModel
title: Models
- isExpanded: false
sections:
diff --git a/docs/source/en/api/models/controlnet_sparsectrl.md b/docs/source/en/api/models/controlnet_sparsectrl.md
new file mode 100644
index 000000000000..d5d7d358c4d2
--- /dev/null
+++ b/docs/source/en/api/models/controlnet_sparsectrl.md
@@ -0,0 +1,46 @@
+
+
+# SparseControlNetModel
+
+SparseControlNetModel is an implementation of ControlNet for [AnimateDiff](https://arxiv.org/abs/2307.04725).
+
+ControlNet was introduced in [Adding Conditional Control to Text-to-Image Diffusion Models](https://huggingface.co/papers/2302.05543) by Lvmin Zhang, Anyi Rao, and Maneesh Agrawala.
+
+The SparseCtrl version of ControlNet was introduced in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion Models](https://arxiv.org/abs/2311.16933) for achieving controlled generation in text-to-video diffusion models by Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai.
+
+The abstract from the paper is:
+
+*The development of text-to-video (T2V), i.e., generating videos with a given text prompt, has been significantly advanced in recent years. However, relying solely on text prompts often results in ambiguous frame composition due to spatial uncertainty. The research community thus leverages the dense structure signals, e.g., per-frame depth/edge sequences, to enhance controllability, whose collection accordingly increases the burden of inference. In this work, we present SparseCtrl to enable flexible structure control with temporally sparse signals, requiring only one or a few inputs, as shown in Figure 1. It incorporates an additional condition encoder to process these sparse signals while leaving the pre-trained T2V model untouched. The proposed approach is compatible with various modalities, including sketches, depth maps, and RGB images, providing more practical control for video generation and promoting applications such as storyboarding, depth rendering, keyframe animation, and interpolation. Extensive experiments demonstrate the generalization of SparseCtrl on both original and personalized T2V generators. Codes and models will be publicly available at [this https URL](https://guoyww.github.io/projects/SparseCtrl).*
+
+## Example for loading SparseControlNetModel
+
+```python
+import torch
+from diffusers import SparseControlNetModel
+
+# fp32 variant in float16
+# 1. Scribble checkpoint
+controlnet = SparseControlNetModel.from_pretrained("guoyww/animatediff-sparsectrl-scribble", torch_dtype=torch.float16)
+
+# 2. RGB checkpoint
+controlnet = SparseControlNetModel.from_pretrained("guoyww/animatediff-sparsectrl-rgb", torch_dtype=torch.float16)
+
+# For loading fp16 variant, pass `variant="fp16"` as an additional parameter
+```
+
+## SparseControlNetModel
+
+[[autodoc]] SparseControlNetModel
+
+## SparseControlNetOutput
+
+[[autodoc]] models.controlnet_sparsectrl.SparseControlNetOutput
diff --git a/docs/source/en/api/pipelines/animatediff.md b/docs/source/en/api/pipelines/animatediff.md
index be18054aed71..835b4d7b7fdd 100644
--- a/docs/source/en/api/pipelines/animatediff.md
+++ b/docs/source/en/api/pipelines/animatediff.md
@@ -100,6 +100,189 @@ AnimateDiff tends to work better with finetuned Stable Diffusion models. If you
+### AnimateDiffSparseControlNetPipeline
+
+[SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion Models](https://arxiv.org/abs/2311.16933) for achieving controlled generation in text-to-video diffusion models by Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai.
+
+The abstract from the paper is:
+
+*The development of text-to-video (T2V), i.e., generating videos with a given text prompt, has been significantly advanced in recent years. However, relying solely on text prompts often results in ambiguous frame composition due to spatial uncertainty. The research community thus leverages the dense structure signals, e.g., per-frame depth/edge sequences, to enhance controllability, whose collection accordingly increases the burden of inference. In this work, we present SparseCtrl to enable flexible structure control with temporally sparse signals, requiring only one or a few inputs, as shown in Figure 1. It incorporates an additional condition encoder to process these sparse signals while leaving the pre-trained T2V model untouched. The proposed approach is compatible with various modalities, including sketches, depth maps, and RGB images, providing more practical control for video generation and promoting applications such as storyboarding, depth rendering, keyframe animation, and interpolation. Extensive experiments demonstrate the generalization of SparseCtrl on both original and personalized T2V generators. Codes and models will be publicly available at [this https URL](https://guoyww.github.io/projects/SparseCtrl).*
+
+SparseCtrl introduces the following checkpoints for controlled text-to-video generation:
+
+- [SparseCtrl Scribble](https://huggingface.co/guoyww/animatediff-sparsectrl-scribble)
+- [SparseCtrl RGB](https://huggingface.co/guoyww/animatediff-sparsectrl-rgb)
+
+#### Using SparseCtrl Scribble
+
+```python
+import torch
+
+from diffusers import AnimateDiffSparseControlNetPipeline
+from diffusers.models import AutoencoderKL, MotionAdapter, SparseControlNetModel
+from diffusers.schedulers import DPMSolverMultistepScheduler
+from diffusers.utils import export_to_gif, load_image
+
+
+model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
+motion_adapter_id = "guoyww/animatediff-motion-adapter-v1-5-3"
+controlnet_id = "guoyww/animatediff-sparsectrl-scribble"
+lora_adapter_id = "guoyww/animatediff-motion-lora-v1-5-3"
+vae_id = "stabilityai/sd-vae-ft-mse"
+device = "cuda"
+
+motion_adapter = MotionAdapter.from_pretrained(motion_adapter_id, torch_dtype=torch.float16).to(device)
+controlnet = SparseControlNetModel.from_pretrained(controlnet_id, torch_dtype=torch.float16).to(device)
+vae = AutoencoderKL.from_pretrained(vae_id, torch_dtype=torch.float16).to(device)
+scheduler = DPMSolverMultistepScheduler.from_pretrained(
+ model_id,
+ subfolder="scheduler",
+ beta_schedule="linear",
+ algorithm_type="dpmsolver++",
+ use_karras_sigmas=True,
+)
+pipe = AnimateDiffSparseControlNetPipeline.from_pretrained(
+ model_id,
+ motion_adapter=motion_adapter,
+ controlnet=controlnet,
+ vae=vae,
+ scheduler=scheduler,
+ torch_dtype=torch.float16,
+).to(device)
+pipe.load_lora_weights(lora_adapter_id, adapter_name="motion_lora")
+pipe.fuse_lora(lora_scale=1.0)
+
+prompt = "an aerial view of a cyberpunk city, night time, neon lights, masterpiece, high quality"
+negative_prompt = "low quality, worst quality, letterboxed"
+
+image_files = [
+ "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-1.png",
+ "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-2.png",
+ "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-3.png"
+]
+condition_frame_indices = [0, 8, 15]
+conditioning_frames = [load_image(img_file) for img_file in image_files]
+
+video = pipe(
+ prompt=prompt,
+ negative_prompt=negative_prompt,
+ num_inference_steps=25,
+ conditioning_frames=conditioning_frames,
+ controlnet_conditioning_scale=1.0,
+ controlnet_frame_indices=condition_frame_indices,
+ generator=torch.Generator().manual_seed(1337),
+).frames[0]
+export_to_gif(video, "output.gif")
+```
+
+Here are some sample outputs:
+
+
+
+
+ an aerial view of a cyberpunk city, night time, neon lights, masterpiece, high quality
+
+
+
+
+
+ 
+
+ |
+
+
+ 
+
+ |
+
+
+ 
+
+ |
+
+
+
+
+ 
+
+ |
+
+
+
+#### Using SparseCtrl RGB
+
+```python
+import torch
+
+from diffusers import AnimateDiffSparseControlNetPipeline
+from diffusers.models import AutoencoderKL, MotionAdapter, SparseControlNetModel
+from diffusers.schedulers import DPMSolverMultistepScheduler
+from diffusers.utils import export_to_gif, load_image
+
+
+model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
+motion_adapter_id = "guoyww/animatediff-motion-adapter-v1-5-3"
+controlnet_id = "guoyww/animatediff-sparsectrl-rgb"
+lora_adapter_id = "guoyww/animatediff-motion-lora-v1-5-3"
+vae_id = "stabilityai/sd-vae-ft-mse"
+device = "cuda"
+
+motion_adapter = MotionAdapter.from_pretrained(motion_adapter_id, torch_dtype=torch.float16).to(device)
+controlnet = SparseControlNetModel.from_pretrained(controlnet_id, torch_dtype=torch.float16).to(device)
+vae = AutoencoderKL.from_pretrained(vae_id, torch_dtype=torch.float16).to(device)
+scheduler = DPMSolverMultistepScheduler.from_pretrained(
+ model_id,
+ subfolder="scheduler",
+ beta_schedule="linear",
+ algorithm_type="dpmsolver++",
+ use_karras_sigmas=True,
+)
+pipe = AnimateDiffSparseControlNetPipeline.from_pretrained(
+ model_id,
+ motion_adapter=motion_adapter,
+ controlnet=controlnet,
+ vae=vae,
+ scheduler=scheduler,
+ torch_dtype=torch.float16,
+).to(device)
+pipe.load_lora_weights(lora_adapter_id, adapter_name="motion_lora")
+
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-firework.png")
+
+video = pipe(
+ prompt="closeup face photo of man in black clothes, night city street, bokeh, fireworks in background",
+ negative_prompt="low quality, worst quality",
+ num_inference_steps=25,
+ conditioning_frames=image,
+ controlnet_frame_indices=[0],
+ controlnet_conditioning_scale=1.0,
+ generator=torch.Generator().manual_seed(42),
+).frames[0]
+export_to_gif(video, "output.gif")
+```
+
+Here are some sample outputs:
+
+
+
+
+ closeup face photo of man in black clothes, night city street, bokeh, fireworks in background
+
+
+
+
+
+ 
+
+ |
+
+
+ 
+
+ |
+
+
+
### AnimateDiffSDXLPipeline
AnimateDiff can also be used with SDXL models. This is currently an experimental feature as only a beta release of the motion adapter checkpoint is available.
@@ -571,7 +754,6 @@ ckpt_path = "https://huggingface.co/Lightricks/LongAnimateDiff/blob/main/lt_long
adapter = MotionAdapter.from_single_file(ckpt_path, torch_dtype=torch.float16)
pipe = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapter=adapter)
-
```
## AnimateDiffPipeline
@@ -580,6 +762,12 @@ pipe = AnimateDiffPipeline.from_pretrained("emilianJR/epiCRealism", motion_adapt
- all
- __call__
+## AnimateDiffSparseControlNetPipeline
+
+[[autodoc]] AnimateDiffSparseControlNetPipeline
+ - all
+ - __call__
+
## AnimateDiffSDXLPipeline
[[autodoc]] AnimateDiffSDXLPipeline
diff --git a/scripts/convert_animatediff_motion_lora_to_diffusers.py b/scripts/convert_animatediff_motion_lora_to_diffusers.py
index c680fdc68462..21567ffa9e7a 100644
--- a/scripts/convert_animatediff_motion_lora_to_diffusers.py
+++ b/scripts/convert_animatediff_motion_lora_to_diffusers.py
@@ -1,6 +1,8 @@
import argparse
+import os
import torch
+from huggingface_hub import create_repo, upload_folder
from safetensors.torch import load_file, save_file
@@ -25,8 +27,14 @@ def convert_motion_module(original_state_dict):
def get_args():
parser = argparse.ArgumentParser()
- parser.add_argument("--ckpt_path", type=str, required=True)
- parser.add_argument("--output_path", type=str, required=True)
+ parser.add_argument("--ckpt_path", type=str, required=True, help="Path to checkpoint")
+ parser.add_argument("--output_path", type=str, required=True, help="Path to output directory")
+ parser.add_argument(
+ "--push_to_hub",
+ action="store_true",
+ default=False,
+ help="Whether to push the converted model to the HF or not",
+ )
return parser.parse_args()
@@ -51,4 +59,11 @@ def get_args():
continue
output_dict.update({f"unet.{module_name}": params})
- save_file(output_dict, f"{args.output_path}/diffusion_pytorch_model.safetensors")
+ os.makedirs(args.output_path, exist_ok=True)
+
+ filepath = os.path.join(args.output_path, "diffusion_pytorch_model.safetensors")
+ save_file(output_dict, filepath)
+
+ if args.push_to_hub:
+ repo_id = create_repo(args.output_path, exist_ok=True).repo_id
+ upload_folder(repo_id=repo_id, folder_path=args.output_path, repo_type="model")
diff --git a/scripts/convert_animatediff_sparsectrl_to_diffusers.py b/scripts/convert_animatediff_sparsectrl_to_diffusers.py
new file mode 100644
index 000000000000..f246dceb97f8
--- /dev/null
+++ b/scripts/convert_animatediff_sparsectrl_to_diffusers.py
@@ -0,0 +1,83 @@
+import argparse
+from typing import Dict
+
+import torch
+import torch.nn as nn
+
+from diffusers import SparseControlNetModel
+
+
+KEYS_RENAME_MAPPING = {
+ ".attention_blocks.0": ".attn1",
+ ".attention_blocks.1": ".attn2",
+ ".attn1.pos_encoder": ".pos_embed",
+ ".ff_norm": ".norm3",
+ ".norms.0": ".norm1",
+ ".norms.1": ".norm2",
+ ".temporal_transformer": "",
+}
+
+
+def convert(original_state_dict: Dict[str, nn.Module]) -> Dict[str, nn.Module]:
+ converted_state_dict = {}
+
+ for key in list(original_state_dict.keys()):
+ renamed_key = key
+ for new_name, old_name in KEYS_RENAME_MAPPING.items():
+ renamed_key = renamed_key.replace(new_name, old_name)
+ converted_state_dict[renamed_key] = original_state_dict.pop(key)
+
+ return converted_state_dict
+
+
+def get_args():
+ parser = argparse.ArgumentParser()
+ parser.add_argument("--ckpt_path", type=str, required=True, help="Path to checkpoint")
+ parser.add_argument("--output_path", type=str, required=True, help="Path to output directory")
+ parser.add_argument(
+ "--max_motion_seq_length",
+ type=int,
+ default=32,
+ help="Max motion sequence length supported by the motion adapter",
+ )
+ parser.add_argument(
+ "--conditioning_channels", type=int, default=4, help="Number of channels in conditioning input to controlnet"
+ )
+ parser.add_argument(
+ "--use_simplified_condition_embedding",
+ action="store_true",
+ default=False,
+ help="Whether or not to use simplified condition embedding. When `conditioning_channels==4` i.e. latent inputs, set this to `True`. When `conditioning_channels==3` i.e. image inputs, set this to `False`",
+ )
+ parser.add_argument(
+ "--save_fp16",
+ action="store_true",
+ default=False,
+ help="Whether or not to save model in fp16 precision along with fp32",
+ )
+ parser.add_argument(
+ "--push_to_hub", action="store_true", default=False, help="Whether or not to push saved model to the HF hub"
+ )
+ return parser.parse_args()
+
+
+if __name__ == "__main__":
+ args = get_args()
+
+ state_dict = torch.load(args.ckpt_path, map_location="cpu")
+ if "state_dict" in state_dict.keys():
+ state_dict: dict = state_dict["state_dict"]
+
+ controlnet = SparseControlNetModel(
+ conditioning_channels=args.conditioning_channels,
+ motion_max_seq_length=args.max_motion_seq_length,
+ use_simplified_condition_embedding=args.use_simplified_condition_embedding,
+ )
+
+ state_dict = convert(state_dict)
+ controlnet.load_state_dict(state_dict, strict=True)
+
+ controlnet.save_pretrained(args.output_path, push_to_hub=args.push_to_hub)
+ if args.save_fp16:
+ controlnet = controlnet.to(dtype=torch.float16)
+ controlnet.save_pretrained(args.output_path, variant="fp16", push_to_hub=args.push_to_hub)
diff --git a/src/diffusers/__init__.py b/src/diffusers/__init__.py
index 6a6607cc376f..f42ccc064624 100644
--- a/src/diffusers/__init__.py
+++ b/src/diffusers/__init__.py
@@ -99,6 +99,7 @@
"SD3ControlNetModel",
"SD3MultiControlNetModel",
"SD3Transformer2DModel",
+ "SparseControlNetModel",
"StableCascadeUNet",
"T2IAdapter",
"T5FilmDecoder",
@@ -231,6 +232,7 @@
"AmusedPipeline",
"AnimateDiffPipeline",
"AnimateDiffSDXLPipeline",
+ "AnimateDiffSparseControlNetPipeline",
"AnimateDiffVideoToVideoPipeline",
"AudioLDM2Pipeline",
"AudioLDM2ProjectionModel",
@@ -533,6 +535,7 @@
SD3ControlNetModel,
SD3MultiControlNetModel,
SD3Transformer2DModel,
+ SparseControlNetModel,
T2IAdapter,
T5FilmDecoder,
Transformer2DModel,
@@ -645,6 +648,7 @@
AmusedPipeline,
AnimateDiffPipeline,
AnimateDiffSDXLPipeline,
+ AnimateDiffSparseControlNetPipeline,
AnimateDiffVideoToVideoPipeline,
AudioLDM2Pipeline,
AudioLDM2ProjectionModel,
diff --git a/src/diffusers/models/__init__.py b/src/diffusers/models/__init__.py
index 39dc149ff6d1..d35786ee7642 100644
--- a/src/diffusers/models/__init__.py
+++ b/src/diffusers/models/__init__.py
@@ -35,6 +35,7 @@
_import_structure["controlnet"] = ["ControlNetModel"]
_import_structure["controlnet_hunyuan"] = ["HunyuanDiT2DControlNetModel", "HunyuanDiT2DMultiControlNetModel"]
_import_structure["controlnet_sd3"] = ["SD3ControlNetModel", "SD3MultiControlNetModel"]
+ _import_structure["controlnet_sparsectrl"] = ["SparseControlNetModel"]
_import_structure["controlnet_xs"] = ["ControlNetXSAdapter", "UNetControlNetXSModel"]
_import_structure["embeddings"] = ["ImageProjection"]
_import_structure["modeling_utils"] = ["ModelMixin"]
@@ -81,6 +82,7 @@
from .controlnet import ControlNetModel
from .controlnet_hunyuan import HunyuanDiT2DControlNetModel, HunyuanDiT2DMultiControlNetModel
from .controlnet_sd3 import SD3ControlNetModel, SD3MultiControlNetModel
+ from .controlnet_sparsectrl import SparseControlNetModel
from .controlnet_xs import ControlNetXSAdapter, UNetControlNetXSModel
from .embeddings import ImageProjection
from .modeling_utils import ModelMixin
diff --git a/src/diffusers/models/controlnet.py b/src/diffusers/models/controlnet.py
index 8fb49d7b547f..d3ae96605077 100644
--- a/src/diffusers/models/controlnet.py
+++ b/src/diffusers/models/controlnet.py
@@ -830,7 +830,6 @@ def forward(
sample = self.mid_block(sample, emb)
# 5. Control net blocks
-
controlnet_down_block_res_samples = ()
for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
diff --git a/src/diffusers/models/controlnet_sparsectrl.py b/src/diffusers/models/controlnet_sparsectrl.py
new file mode 100644
index 000000000000..bc1273aaab7d
--- /dev/null
+++ b/src/diffusers/models/controlnet_sparsectrl.py
@@ -0,0 +1,791 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from ..configuration_utils import ConfigMixin, register_to_config
+from ..utils import BaseOutput, logging
+from .attention_processor import (
+ ADDED_KV_ATTENTION_PROCESSORS,
+ CROSS_ATTENTION_PROCESSORS,
+ AttentionProcessor,
+ AttnAddedKVProcessor,
+ AttnProcessor,
+)
+from .embeddings import TimestepEmbedding, Timesteps
+from .modeling_utils import ModelMixin
+from .unets.unet_2d_blocks import UNetMidBlock2DCrossAttn
+from .unets.unet_2d_condition import UNet2DConditionModel
+from .unets.unet_3d_blocks import (
+ CrossAttnDownBlockMotion,
+ DownBlockMotion,
+)
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+
+@dataclass
+class SparseControlNetOutput(BaseOutput):
+ """
+ The output of [`SparseControlNetModel`].
+
+ 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 middle 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 SparseControlNetConditioningEmbedding(nn.Module):
+ 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: torch.Tensor) -> torch.Tensor:
+ 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 SparseControlNetModel(ModelMixin, ConfigMixin):
+ """
+ A SparseControlNet model as described in [SparseCtrl: Adding Sparse Controls to Text-to-Video Diffusion
+ Models](https://arxiv.org/abs/2311.16933).
+
+ Args:
+ in_channels (`int`, defaults to 4):
+ The number of channels in the input sample.
+ conditioning_channels (`int`, defaults to 4):
+ The number of input channels in the controlnet conditional embedding module. If
+ `concat_condition_embedding` is True, the value provided here is incremented by 1.
+ 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`].
+ transformer_layers_per_mid_block (`int` or `Tuple[int]`, *optional*, defaults to 1):
+ The number of transformer layers to use in each layer in the middle block.
+ attention_head_dim (`int` or `Tuple[int]`, defaults to 8):
+ The dimension of the attention heads.
+ num_attention_heads (`int` or `Tuple[int]`, *optional*):
+ The number of heads to use for multi-head attention.
+ use_linear_projection (`bool`, defaults to `False`):
+ 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`.
+ conditioning_embedding_out_channels (`Tuple[int]`, 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
+ controlnet_conditioning_channel_order (`str`, defaults to `rgb`):
+ motion_max_seq_length (`int`, defaults to `32`):
+ The maximum sequence length to use in the motion module.
+ motion_num_attention_heads (`int` or `Tuple[int]`, defaults to `8`):
+ The number of heads to use in each attention layer of the motion module.
+ concat_conditioning_mask (`bool`, defaults to `True`):
+ use_simplified_condition_embedding (`bool`, defaults to `True`):
+ """
+
+ _supports_gradient_checkpointing = True
+
+ @register_to_config
+ def __init__(
+ self,
+ in_channels: int = 4,
+ conditioning_channels: int = 4,
+ flip_sin_to_cos: bool = True,
+ freq_shift: int = 0,
+ down_block_types: Tuple[str, ...] = (
+ "CrossAttnDownBlockMotion",
+ "CrossAttnDownBlockMotion",
+ "CrossAttnDownBlockMotion",
+ "DownBlockMotion",
+ ),
+ 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 = 768,
+ transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ transformer_layers_per_mid_block: Optional[Union[int, Tuple[int]]] = None,
+ temporal_transformer_layers_per_block: Union[int, Tuple[int, ...]] = 1,
+ attention_head_dim: Union[int, Tuple[int, ...]] = 8,
+ num_attention_heads: Optional[Union[int, Tuple[int, ...]]] = None,
+ use_linear_projection: bool = False,
+ upcast_attention: bool = False,
+ resnet_time_scale_shift: str = "default",
+ conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
+ global_pool_conditions: bool = False,
+ controlnet_conditioning_channel_order: str = "rgb",
+ motion_max_seq_length: int = 32,
+ motion_num_attention_heads: int = 8,
+ concat_conditioning_mask: bool = True,
+ use_simplified_condition_embedding: bool = True,
+ ):
+ super().__init__()
+ self.use_simplified_condition_embedding = use_simplified_condition_embedding
+
+ # If `num_attention_heads` is not defined (which is the case for most models)
+ # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+ # The reason for this behavior is to correct for incorrectly named variables that were introduced
+ # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+ # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+ # which is why we correct for the naming here.
+ num_attention_heads = num_attention_heads or attention_head_dim
+
+ # Check inputs
+ if len(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)
+ if isinstance(temporal_transformer_layers_per_block, int):
+ temporal_transformer_layers_per_block = [temporal_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
+ )
+
+ if concat_conditioning_mask:
+ conditioning_channels = conditioning_channels + 1
+
+ self.concat_conditioning_mask = concat_conditioning_mask
+
+ # control net conditioning embedding
+ if use_simplified_condition_embedding:
+ self.controlnet_cond_embedding = zero_module(
+ nn.Conv2d(conditioning_channels, block_out_channels[0], kernel_size=3, padding=1)
+ )
+ else:
+ self.controlnet_cond_embedding = SparseControlNetConditioningEmbedding(
+ conditioning_embedding_channels=block_out_channels[0],
+ block_out_channels=conditioning_embedding_out_channels,
+ conditioning_channels=conditioning_channels,
+ )
+
+ # 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,
+ )
+
+ self.down_blocks = nn.ModuleList([])
+ self.controlnet_down_blocks = nn.ModuleList([])
+
+ if isinstance(cross_attention_dim, int):
+ cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+ 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)
+
+ if isinstance(motion_num_attention_heads, int):
+ motion_num_attention_heads = (motion_num_attention_heads,) * len(down_block_types)
+
+ # down
+ output_channel = block_out_channels[0]
+
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ 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
+
+ if down_block_type == "CrossAttnDownBlockMotion":
+ down_block = CrossAttnDownBlockMotion(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ dropout=0,
+ num_layers=layers_per_block,
+ transformer_layers_per_block=transformer_layers_per_block[i],
+ resnet_eps=norm_eps,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ resnet_pre_norm=True,
+ num_attention_heads=num_attention_heads[i],
+ cross_attention_dim=cross_attention_dim[i],
+ add_downsample=not is_final_block,
+ dual_cross_attention=False,
+ use_linear_projection=use_linear_projection,
+ only_cross_attention=only_cross_attention[i],
+ upcast_attention=upcast_attention,
+ temporal_num_attention_heads=motion_num_attention_heads[i],
+ temporal_max_seq_length=motion_max_seq_length,
+ temporal_transformer_layers_per_block=temporal_transformer_layers_per_block[i],
+ temporal_double_self_attention=False,
+ )
+ elif down_block_type == "DownBlockMotion":
+ down_block = DownBlockMotion(
+ in_channels=input_channel,
+ out_channels=output_channel,
+ temb_channels=time_embed_dim,
+ dropout=0,
+ num_layers=layers_per_block,
+ resnet_eps=norm_eps,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ resnet_pre_norm=True,
+ add_downsample=not is_final_block,
+ temporal_num_attention_heads=motion_num_attention_heads[i],
+ temporal_max_seq_length=motion_max_seq_length,
+ temporal_double_self_attention=False,
+ temporal_transformer_layers_per_block=temporal_transformer_layers_per_block[i],
+ )
+ else:
+ raise ValueError(
+ "Invalid `block_type` encountered. Must be one of `CrossAttnDownBlockMotion` or `DownBlockMotion`"
+ )
+
+ self.down_blocks.append(down_block)
+
+ for _ in range(layers_per_block):
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_down_blocks.append(controlnet_block)
+
+ if not is_final_block:
+ controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_down_blocks.append(controlnet_block)
+
+ # mid
+ mid_block_channels = block_out_channels[-1]
+
+ controlnet_block = nn.Conv2d(mid_block_channels, mid_block_channels, kernel_size=1)
+ controlnet_block = zero_module(controlnet_block)
+ self.controlnet_mid_block = controlnet_block
+
+ if transformer_layers_per_mid_block is None:
+ transformer_layers_per_mid_block = (
+ transformer_layers_per_block[-1] if isinstance(transformer_layers_per_block[-1], int) else 1
+ )
+
+ self.mid_block = UNetMidBlock2DCrossAttn(
+ in_channels=mid_block_channels,
+ temb_channels=time_embed_dim,
+ dropout=0,
+ num_layers=1,
+ transformer_layers_per_block=transformer_layers_per_mid_block,
+ resnet_eps=norm_eps,
+ resnet_time_scale_shift=resnet_time_scale_shift,
+ resnet_act_fn=act_fn,
+ resnet_groups=norm_num_groups,
+ resnet_pre_norm=True,
+ num_attention_heads=num_attention_heads[-1],
+ output_scale_factor=mid_block_scale_factor,
+ cross_attention_dim=cross_attention_dim[-1],
+ dual_cross_attention=False,
+ use_linear_projection=use_linear_projection,
+ upcast_attention=upcast_attention,
+ attention_type="default",
+ )
+
+ @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,
+ ) -> "SparseControlNetModel":
+ r"""
+ Instantiate a [`SparseControlNetModel`] from [`UNet2DConditionModel`].
+
+ Parameters:
+ unet (`UNet2DConditionModel`):
+ The UNet model weights to copy to the [`SparseControlNetModel`]. 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
+ )
+ down_block_types = unet.config.down_block_types
+
+ for i in range(len(down_block_types)):
+ if "CrossAttn" in down_block_types[i]:
+ down_block_types[i] = "CrossAttnDownBlockMotion"
+ elif "Down" in down_block_types[i]:
+ down_block_types[i] = "DownBlockMotion"
+ else:
+ raise ValueError("Invalid `block_type` encountered. Must be a cross-attention or down block")
+
+ controlnet = cls(
+ in_channels=unet.config.in_channels,
+ conditioning_channels=conditioning_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,
+ transformer_layers_per_block=transformer_layers_per_block,
+ attention_head_dim=unet.config.attention_head_dim,
+ num_attention_heads=unet.config.num_attention_heads,
+ use_linear_projection=unet.config.use_linear_projection,
+ upcast_attention=unet.config.upcast_attention,
+ resnet_time_scale_shift=unet.config.resnet_time_scale_shift,
+ conditioning_embedding_out_channels=conditioning_embedding_out_channels,
+ controlnet_conditioning_channel_order=controlnet_conditioning_channel_order,
+ )
+
+ if load_weights_from_unet:
+ controlnet.conv_in.load_state_dict(unet.conv_in.state_dict(), strict=False)
+ controlnet.time_proj.load_state_dict(unet.time_proj.state_dict(), strict=False)
+ controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict(), strict=False)
+ controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict(), strict=False)
+ controlnet.mid_block.load_state_dict(unet.mid_block.state_dict(), strict=False)
+
+ 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()
+
+ 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 _set_gradient_checkpointing(self, module, value: bool = False) -> None:
+ if isinstance(module, (CrossAttnDownBlockMotion, DownBlockMotion, UNetMidBlock2DCrossAttn)):
+ module.gradient_checkpointing = value
+
+ def forward(
+ self,
+ sample: torch.Tensor,
+ timestep: Union[torch.Tensor, float, int],
+ encoder_hidden_states: torch.Tensor,
+ controlnet_cond: torch.Tensor,
+ conditioning_scale: float = 1.0,
+ timestep_cond: Optional[torch.Tensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ conditioning_mask: Optional[torch.Tensor] = None,
+ guess_mode: bool = False,
+ return_dict: bool = True,
+ ) -> Union[SparseControlNetOutput, Tuple[Tuple[torch.Tensor, ...], torch.Tensor]]:
+ """
+ The [`SparseControlNetModel`] forward method.
+
+ Args:
+ sample (`torch.Tensor`):
+ 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.Tensor`):
+ 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`.
+ guess_mode (`bool`, defaults to `False`):
+ In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
+ you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
+ 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.
+ """
+ sample_batch_size, sample_channels, sample_num_frames, sample_height, sample_width = sample.shape
+ sample = torch.zeros_like(sample)
+
+ # check channel order
+ channel_order = self.config.controlnet_conditioning_channel_order
+
+ if channel_order == "rgb":
+ # in rgb order by default
+ ...
+ elif channel_order == "bgr":
+ controlnet_cond = torch.flip(controlnet_cond, dims=[1])
+ 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):
+ # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+ # This would be a good case for the `match` statement (Python 3.10+)
+ is_mps = sample.device.type == "mps"
+ if isinstance(timestep, float):
+ dtype = torch.float32 if is_mps else torch.float64
+ else:
+ dtype = torch.int32 if is_mps else torch.int64
+ timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+ elif len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ t_emb = self.time_proj(timesteps)
+
+ # 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)
+ emb = emb.repeat_interleave(sample_num_frames, dim=0)
+ encoder_hidden_states = encoder_hidden_states.repeat_interleave(sample_num_frames, dim=0)
+
+ # 2. pre-process
+ batch_size, channels, num_frames, height, width = sample.shape
+
+ sample = sample.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+ sample = self.conv_in(sample)
+
+ batch_frames, channels, height, width = sample.shape
+ sample = sample[:, None].reshape(sample_batch_size, sample_num_frames, channels, height, width)
+
+ if self.concat_conditioning_mask:
+ controlnet_cond = torch.cat([controlnet_cond, conditioning_mask], dim=1)
+
+ batch_size, channels, num_frames, height, width = controlnet_cond.shape
+ controlnet_cond = controlnet_cond.permute(0, 2, 1, 3, 4).reshape(
+ batch_size * num_frames, channels, height, width
+ )
+ controlnet_cond = self.controlnet_cond_embedding(controlnet_cond)
+ batch_frames, channels, height, width = controlnet_cond.shape
+ controlnet_cond = controlnet_cond[:, None].reshape(batch_size, num_frames, channels, height, width)
+
+ sample = sample + controlnet_cond
+
+ batch_size, num_frames, channels, height, width = sample.shape
+ sample = sample.reshape(sample_batch_size * sample_num_frames, channels, height, width)
+
+ # 3. down
+ down_block_res_samples = (sample,)
+ for downsample_block in self.down_blocks:
+ if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+ sample, res_samples = downsample_block(
+ hidden_states=sample,
+ temb=emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ num_frames=num_frames,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample, res_samples = downsample_block(hidden_states=sample, temb=emb, num_frames=num_frames)
+
+ down_block_res_samples += res_samples
+
+ # 4. mid
+ if self.mid_block is not None:
+ if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+ sample = self.mid_block(
+ sample,
+ emb,
+ encoder_hidden_states=encoder_hidden_states,
+ attention_mask=attention_mask,
+ cross_attention_kwargs=cross_attention_kwargs,
+ )
+ else:
+ sample = self.mid_block(sample, emb)
+
+ # 5. Control net blocks
+ controlnet_down_block_res_samples = ()
+
+ for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
+ down_block_res_sample = controlnet_block(down_block_res_sample)
+ controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
+
+ down_block_res_samples = controlnet_down_block_res_samples
+ mid_block_res_sample = self.controlnet_mid_block(sample)
+
+ # 6. scaling
+ if guess_mode and not self.config.global_pool_conditions:
+ scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device) # 0.1 to 1.0
+ scales = scales * conditioning_scale
+ down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
+ mid_block_res_sample = mid_block_res_sample * scales[-1] # last one
+ else:
+ 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 SparseControlNetOutput(
+ down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
+ )
+
+
+# Copied from diffusers.models.controlnet.zero_module
+def zero_module(module: nn.Module) -> nn.Module:
+ for p in module.parameters():
+ nn.init.zeros_(p)
+ return module
diff --git a/src/diffusers/models/unets/unet_3d_blocks.py b/src/diffusers/models/unets/unet_3d_blocks.py
index 14f2df0b582c..51c743a14d40 100644
--- a/src/diffusers/models/unets/unet_3d_blocks.py
+++ b/src/diffusers/models/unets/unet_3d_blocks.py
@@ -966,6 +966,7 @@ def __init__(
temporal_num_attention_heads: Union[int, Tuple[int]] = 1,
temporal_cross_attention_dim: Optional[int] = None,
temporal_max_seq_length: int = 32,
+ temporal_double_self_attention: bool = True,
temporal_transformer_layers_per_block: Union[int, Tuple[int]] = 1,
):
super().__init__()
@@ -1016,6 +1017,7 @@ def __init__(
positional_embeddings="sinusoidal",
num_positional_embeddings=temporal_max_seq_length,
attention_head_dim=out_channels // temporal_num_attention_heads[i],
+ double_self_attention=temporal_double_self_attention,
)
)
@@ -1118,6 +1120,7 @@ def __init__(
temporal_num_attention_heads: int = 8,
temporal_max_seq_length: int = 32,
temporal_transformer_layers_per_block: Union[int, Tuple[int]] = 1,
+ temporal_double_self_attention: bool = True,
):
super().__init__()
resnets = []
@@ -1199,6 +1202,7 @@ def __init__(
positional_embeddings="sinusoidal",
num_positional_embeddings=temporal_max_seq_length,
attention_head_dim=out_channels // temporal_num_attention_heads,
+ double_self_attention=temporal_double_self_attention,
)
)
diff --git a/src/diffusers/pipelines/__init__.py b/src/diffusers/pipelines/__init__.py
index 1d5fd5c2d094..7bc50b297566 100644
--- a/src/diffusers/pipelines/__init__.py
+++ b/src/diffusers/pipelines/__init__.py
@@ -119,6 +119,7 @@
_import_structure["animatediff"] = [
"AnimateDiffPipeline",
"AnimateDiffSDXLPipeline",
+ "AnimateDiffSparseControlNetPipeline",
"AnimateDiffVideoToVideoPipeline",
]
_import_structure["audioldm"] = ["AudioLDMPipeline"]
@@ -413,7 +414,12 @@
from ..utils.dummy_torch_and_transformers_objects import *
else:
from .amused import AmusedImg2ImgPipeline, AmusedInpaintPipeline, AmusedPipeline
- from .animatediff import AnimateDiffPipeline, AnimateDiffSDXLPipeline, AnimateDiffVideoToVideoPipeline
+ from .animatediff import (
+ AnimateDiffPipeline,
+ AnimateDiffSDXLPipeline,
+ AnimateDiffSparseControlNetPipeline,
+ AnimateDiffVideoToVideoPipeline,
+ )
from .audioldm import AudioLDMPipeline
from .audioldm2 import (
AudioLDM2Pipeline,
diff --git a/src/diffusers/pipelines/animatediff/__init__.py b/src/diffusers/pipelines/animatediff/__init__.py
index ae6b67b1924c..5f4d6f29391c 100644
--- a/src/diffusers/pipelines/animatediff/__init__.py
+++ b/src/diffusers/pipelines/animatediff/__init__.py
@@ -23,6 +23,7 @@
else:
_import_structure["pipeline_animatediff"] = ["AnimateDiffPipeline"]
_import_structure["pipeline_animatediff_sdxl"] = ["AnimateDiffSDXLPipeline"]
+ _import_structure["pipeline_animatediff_sparsectrl"] = ["AnimateDiffSparseControlNetPipeline"]
_import_structure["pipeline_animatediff_video2video"] = ["AnimateDiffVideoToVideoPipeline"]
if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
@@ -35,6 +36,7 @@
else:
from .pipeline_animatediff import AnimateDiffPipeline
from .pipeline_animatediff_sdxl import AnimateDiffSDXLPipeline
+ from .pipeline_animatediff_sparsectrl import AnimateDiffSparseControlNetPipeline
from .pipeline_animatediff_video2video import AnimateDiffVideoToVideoPipeline
from .pipeline_output import AnimateDiffPipelineOutput
diff --git a/src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py b/src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py
new file mode 100644
index 000000000000..e9e0d518c806
--- /dev/null
+++ b/src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py
@@ -0,0 +1,1008 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import numpy as np
+import PIL
+import torch
+import torch.nn.functional as F
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import IPAdapterMixin, StableDiffusionLoraLoaderMixin, TextualInversionLoaderMixin
+from ...models import AutoencoderKL, ImageProjection, UNet2DConditionModel, UNetMotionModel
+from ...models.controlnet_sparsectrl import SparseControlNetModel
+from ...models.lora import adjust_lora_scale_text_encoder
+from ...models.unets.unet_motion_model import MotionAdapter
+from ...schedulers import KarrasDiffusionSchedulers
+from ...utils import (
+ USE_PEFT_BACKEND,
+ logging,
+ replace_example_docstring,
+ scale_lora_layers,
+ unscale_lora_layers,
+)
+from ...utils.torch_utils import is_compiled_module, randn_tensor
+from ...video_processor import VideoProcessor
+from ..free_init_utils import FreeInitMixin
+from ..pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from .pipeline_output import AnimateDiffPipelineOutput
+
+
+logger = logging.get_logger(__name__) # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+ Examples:
+ ```python
+ >>> import torch
+ >>> from diffusers import AnimateDiffSparseControlNetPipeline
+ >>> from diffusers.models import AutoencoderKL, MotionAdapter, SparseControlNetModel
+ >>> from diffusers.schedulers import DPMSolverMultistepScheduler
+ >>> from diffusers.utils import export_to_gif, load_image
+
+ >>> model_id = "SG161222/Realistic_Vision_V5.1_noVAE"
+ >>> motion_adapter_id = "guoyww/animatediff-motion-adapter-v1-5-3"
+ >>> controlnet_id = "guoyww/animatediff-sparsectrl-scribble"
+ >>> lora_adapter_id = "guoyww/animatediff-motion-lora-v1-5-3"
+ >>> vae_id = "stabilityai/sd-vae-ft-mse"
+ >>> device = "cuda"
+
+ >>> motion_adapter = MotionAdapter.from_pretrained(motion_adapter_id, torch_dtype=torch.float16).to(device)
+ >>> controlnet = SparseControlNetModel.from_pretrained(controlnet_id, torch_dtype=torch.float16).to(device)
+ >>> vae = AutoencoderKL.from_pretrained(vae_id, torch_dtype=torch.float16).to(device)
+ >>> scheduler = DPMSolverMultistepScheduler.from_pretrained(
+ ... model_id,
+ ... subfolder="scheduler",
+ ... beta_schedule="linear",
+ ... algorithm_type="dpmsolver++",
+ ... use_karras_sigmas=True,
+ ... )
+ >>> pipe = AnimateDiffSparseControlNetPipeline.from_pretrained(
+ ... model_id,
+ ... motion_adapter=motion_adapter,
+ ... controlnet=controlnet,
+ ... vae=vae,
+ ... scheduler=scheduler,
+ ... torch_dtype=torch.float16,
+ ... ).to(device)
+ >>> pipe.load_lora_weights(lora_adapter_id, adapter_name="motion_lora")
+ >>> pipe.fuse_lora(lora_scale=1.0)
+
+ >>> prompt = "an aerial view of a cyberpunk city, night time, neon lights, masterpiece, high quality"
+ >>> negative_prompt = "low quality, worst quality, letterboxed"
+
+ >>> image_files = [
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-1.png",
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-2.png",
+ ... "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-3.png",
+ ... ]
+ >>> condition_frame_indices = [0, 8, 15]
+ >>> conditioning_frames = [load_image(img_file) for img_file in image_files]
+
+ >>> video = pipe(
+ ... prompt=prompt,
+ ... negative_prompt=negative_prompt,
+ ... num_inference_steps=25,
+ ... conditioning_frames=conditioning_frames,
+ ... controlnet_conditioning_scale=1.0,
+ ... controlnet_frame_indices=condition_frame_indices,
+ ... generator=torch.Generator().manual_seed(1337),
+ ... ).frames[0]
+ >>> export_to_gif(video, "output.gif")
+ ```
+"""
+
+
+# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
+def retrieve_latents(
+ encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
+):
+ if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
+ return encoder_output.latent_dist.sample(generator)
+ elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
+ return encoder_output.latent_dist.mode()
+ elif hasattr(encoder_output, "latents"):
+ return encoder_output.latents
+ else:
+ raise AttributeError("Could not access latents of provided encoder_output")
+
+
+class AnimateDiffSparseControlNetPipeline(
+ DiffusionPipeline,
+ StableDiffusionMixin,
+ TextualInversionLoaderMixin,
+ IPAdapterMixin,
+ StableDiffusionLoraLoaderMixin,
+ FreeInitMixin,
+):
+ r"""
+ Pipeline for controlled text-to-video generation using the method described in [SparseCtrl: Adding Sparse Controls
+ to Text-to-Video Diffusion Models](https://arxiv.org/abs/2311.16933).
+
+ This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+ implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+ The pipeline also inherits the following loading methods:
+ - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+ - [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+ - [`~loaders.StableDiffusionLoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+ - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+ Args:
+ vae ([`AutoencoderKL`]):
+ Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+ text_encoder ([`CLIPTextModel`]):
+ Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+ tokenizer (`CLIPTokenizer`):
+ A [`~transformers.CLIPTokenizer`] to tokenize text.
+ unet ([`UNet2DConditionModel`]):
+ A [`UNet2DConditionModel`] used to create a UNetMotionModel to denoise the encoded video latents.
+ motion_adapter ([`MotionAdapter`]):
+ A [`MotionAdapter`] to be used in combination with `unet` to denoise the encoded video latents.
+ scheduler ([`SchedulerMixin`]):
+ A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+ [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+ """
+
+ model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+ _optional_components = ["feature_extractor", "image_encoder", "motion_adapter"]
+ _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+ def __init__(
+ self,
+ vae: AutoencoderKL,
+ text_encoder: CLIPTextModel,
+ tokenizer: CLIPTokenizer,
+ unet: Union[UNet2DConditionModel, UNetMotionModel],
+ motion_adapter: MotionAdapter,
+ controlnet: SparseControlNetModel,
+ scheduler: KarrasDiffusionSchedulers,
+ feature_extractor: CLIPImageProcessor = None,
+ image_encoder: CLIPVisionModelWithProjection = None,
+ ):
+ super().__init__()
+ if isinstance(unet, UNet2DConditionModel):
+ unet = UNetMotionModel.from_unet2d(unet, motion_adapter)
+
+ self.register_modules(
+ vae=vae,
+ text_encoder=text_encoder,
+ tokenizer=tokenizer,
+ unet=unet,
+ motion_adapter=motion_adapter,
+ controlnet=controlnet,
+ scheduler=scheduler,
+ feature_extractor=feature_extractor,
+ image_encoder=image_encoder,
+ )
+ self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+ self.video_processor = VideoProcessor(do_resize=False, vae_scale_factor=self.vae_scale_factor)
+ self.control_image_processor = VaeImageProcessor(
+ vae_scale_factor=self.vae_scale_factor, do_convert_rgb=True, do_normalize=False
+ )
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_prompt with num_images_per_prompt -> num_videos_per_prompt
+ def encode_prompt(
+ self,
+ prompt,
+ device,
+ num_images_per_prompt,
+ do_classifier_free_guidance,
+ negative_prompt=None,
+ prompt_embeds: Optional[torch.Tensor] = None,
+ negative_prompt_embeds: Optional[torch.Tensor] = None,
+ lora_scale: Optional[float] = None,
+ clip_skip: Optional[int] = None,
+ ):
+ r"""
+ Encodes the prompt into text encoder hidden states.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ prompt to be encoded
+ device: (`torch.device`):
+ torch device
+ num_images_per_prompt (`int`):
+ number of images that should be generated per prompt
+ do_classifier_free_guidance (`bool`):
+ whether to use classifier free guidance or not
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts not to guide the image generation. If not defined, one has to pass
+ `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
+ less than `1`).
+ prompt_embeds (`torch.Tensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+ provided, text embeddings will be generated from `prompt` input argument.
+ negative_prompt_embeds (`torch.Tensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+ weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+ argument.
+ lora_scale (`float`, *optional*):
+ A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ """
+ # set lora scale so that monkey patched LoRA
+ # function of text encoder can correctly access it
+ if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
+ self._lora_scale = lora_scale
+
+ # dynamically adjust the LoRA scale
+ if not USE_PEFT_BACKEND:
+ adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+ else:
+ scale_lora_layers(self.text_encoder, lora_scale)
+
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ if prompt_embeds is None:
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+ text_inputs = self.tokenizer(
+ prompt,
+ padding="max_length",
+ max_length=self.tokenizer.model_max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+ text_input_ids = text_inputs.input_ids
+ untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+ if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+ text_input_ids, untruncated_ids
+ ):
+ removed_text = self.tokenizer.batch_decode(
+ untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+ )
+ logger.warning(
+ "The following part of your input was truncated because CLIP can only handle sequences up to"
+ f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = text_inputs.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ if clip_skip is None:
+ prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+ prompt_embeds = prompt_embeds[0]
+ else:
+ prompt_embeds = self.text_encoder(
+ text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+ )
+ # Access the `hidden_states` first, that contains a tuple of
+ # all the hidden states from the encoder layers. Then index into
+ # the tuple to access the hidden states from the desired layer.
+ prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+ # We also need to apply the final LayerNorm here to not mess with the
+ # representations. The `last_hidden_states` that we typically use for
+ # obtaining the final prompt representations passes through the LayerNorm
+ # layer.
+ prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+ if self.text_encoder is not None:
+ prompt_embeds_dtype = self.text_encoder.dtype
+ elif self.unet is not None:
+ prompt_embeds_dtype = self.unet.dtype
+ else:
+ prompt_embeds_dtype = prompt_embeds.dtype
+
+ prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ bs_embed, seq_len, _ = prompt_embeds.shape
+ # duplicate text embeddings for each generation per prompt, using mps friendly method
+ prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+ # get unconditional embeddings for classifier free guidance
+ if do_classifier_free_guidance and negative_prompt_embeds is None:
+ uncond_tokens: List[str]
+ if negative_prompt is None:
+ uncond_tokens = [""] * batch_size
+ elif prompt is not None and type(prompt) is not type(negative_prompt):
+ raise TypeError(
+ f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+ f" {type(prompt)}."
+ )
+ elif isinstance(negative_prompt, str):
+ uncond_tokens = [negative_prompt]
+ elif batch_size != len(negative_prompt):
+ raise ValueError(
+ f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+ f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+ " the batch size of `prompt`."
+ )
+ else:
+ uncond_tokens = negative_prompt
+
+ # textual inversion: process multi-vector tokens if necessary
+ if isinstance(self, TextualInversionLoaderMixin):
+ uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+ max_length = prompt_embeds.shape[1]
+ uncond_input = self.tokenizer(
+ uncond_tokens,
+ padding="max_length",
+ max_length=max_length,
+ truncation=True,
+ return_tensors="pt",
+ )
+
+ if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+ attention_mask = uncond_input.attention_mask.to(device)
+ else:
+ attention_mask = None
+
+ negative_prompt_embeds = self.text_encoder(
+ uncond_input.input_ids.to(device),
+ attention_mask=attention_mask,
+ )
+ negative_prompt_embeds = negative_prompt_embeds[0]
+
+ if do_classifier_free_guidance:
+ # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+ seq_len = negative_prompt_embeds.shape[1]
+
+ negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+ negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+ negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+ if self.text_encoder is not None:
+ if isinstance(self, StableDiffusionLoraLoaderMixin) and USE_PEFT_BACKEND:
+ # Retrieve the original scale by scaling back the LoRA layers
+ unscale_lora_layers(self.text_encoder, lora_scale)
+
+ return prompt_embeds, negative_prompt_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.encode_image
+ def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+ dtype = next(self.image_encoder.parameters()).dtype
+
+ if not isinstance(image, torch.Tensor):
+ image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+ image = image.to(device=device, dtype=dtype)
+ if output_hidden_states:
+ image_enc_hidden_states = self.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)
+ uncond_image_enc_hidden_states = self.image_encoder(
+ torch.zeros_like(image), output_hidden_states=True
+ ).hidden_states[-2]
+ uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+ num_images_per_prompt, dim=0
+ )
+ return image_enc_hidden_states, uncond_image_enc_hidden_states
+ else:
+ image_embeds = self.image_encoder(image).image_embeds
+ image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+ uncond_image_embeds = torch.zeros_like(image_embeds)
+
+ return image_embeds, uncond_image_embeds
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_ip_adapter_image_embeds
+ def prepare_ip_adapter_image_embeds(
+ self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+ ):
+ image_embeds = []
+ if do_classifier_free_guidance:
+ negative_image_embeds = []
+ 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(self.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(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+ )
+
+ for single_ip_adapter_image, image_proj_layer in zip(
+ ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+ ):
+ output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+ single_image_embeds, single_negative_image_embeds = self.encode_image(
+ single_ip_adapter_image, device, 1, output_hidden_state
+ )
+
+ image_embeds.append(single_image_embeds[None, :])
+ if do_classifier_free_guidance:
+ negative_image_embeds.append(single_negative_image_embeds[None, :])
+ else:
+ 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)
+ negative_image_embeds.append(single_negative_image_embeds)
+ image_embeds.append(single_image_embeds)
+
+ ip_adapter_image_embeds = []
+ for i, single_image_embeds in enumerate(image_embeds):
+ single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
+ if do_classifier_free_guidance:
+ single_negative_image_embeds = torch.cat([negative_image_embeds[i]] * num_images_per_prompt, dim=0)
+ single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds], dim=0)
+
+ single_image_embeds = single_image_embeds.to(device=device)
+ ip_adapter_image_embeds.append(single_image_embeds)
+
+ return ip_adapter_image_embeds
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis/pipeline_text_to_video_synth.TextToVideoSDPipeline.decode_latents
+ def decode_latents(self, latents):
+ latents = 1 / self.vae.config.scaling_factor * latents
+
+ batch_size, channels, num_frames, height, width = latents.shape
+ latents = latents.permute(0, 2, 1, 3, 4).reshape(batch_size * num_frames, channels, height, width)
+
+ image = self.vae.decode(latents).sample
+ video = image[None, :].reshape((batch_size, num_frames, -1) + image.shape[2:]).permute(0, 2, 1, 3, 4)
+ # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+ video = video.float()
+ return video
+
+ # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+ def prepare_extra_step_kwargs(self, generator, eta):
+ # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+ # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+ # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+ # and should be between [0, 1]
+
+ accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ extra_step_kwargs = {}
+ if accepts_eta:
+ extra_step_kwargs["eta"] = eta
+
+ # check if the scheduler accepts generator
+ accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+ if accepts_generator:
+ extra_step_kwargs["generator"] = generator
+ return extra_step_kwargs
+
+ def check_inputs(
+ self,
+ prompt,
+ height,
+ width,
+ negative_prompt=None,
+ prompt_embeds=None,
+ negative_prompt_embeds=None,
+ ip_adapter_image=None,
+ ip_adapter_image_embeds=None,
+ callback_on_step_end_tensor_inputs=None,
+ image=None,
+ controlnet_conditioning_scale: float = 1.0,
+ ):
+ if height % 8 != 0 or width % 8 != 0:
+ raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+ if callback_on_step_end_tensor_inputs is not None and not all(
+ k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+ ):
+ raise ValueError(
+ f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+ )
+
+ if prompt is not None and prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+ " only forward one of the two."
+ )
+ elif prompt is None and prompt_embeds is None:
+ raise ValueError(
+ "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+ )
+ elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+ raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+ if negative_prompt is not None and negative_prompt_embeds is not None:
+ raise ValueError(
+ f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+ f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+ )
+
+ if prompt_embeds is not None and negative_prompt_embeds is not None:
+ if prompt_embeds.shape != negative_prompt_embeds.shape:
+ raise ValueError(
+ "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+ f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+ f" {negative_prompt_embeds.shape}."
+ )
+
+ if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+ raise ValueError(
+ "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+ )
+
+ if ip_adapter_image_embeds is not None:
+ if not isinstance(ip_adapter_image_embeds, list):
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+ )
+ elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+ raise ValueError(
+ f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+ )
+
+ is_compiled = hasattr(F, "scaled_dot_product_attention") and isinstance(
+ self.controlnet, torch._dynamo.eval_frame.OptimizedModule
+ )
+
+ # check `image`
+ if (
+ isinstance(self.controlnet, SparseControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, SparseControlNetModel)
+ ):
+ if isinstance(image, list):
+ for image_ in image:
+ self.check_image(image_, prompt, prompt_embeds)
+ else:
+ self.check_image(image, prompt, prompt_embeds)
+ else:
+ assert False
+
+ # Check `controlnet_conditioning_scale`
+ if (
+ isinstance(self.controlnet, SparseControlNetModel)
+ or is_compiled
+ and isinstance(self.controlnet._orig_mod, SparseControlNetModel)
+ ):
+ if not isinstance(controlnet_conditioning_scale, float):
+ raise TypeError("For single controlnet: `controlnet_conditioning_scale` must be type `float`.")
+ else:
+ assert False
+
+ # Copied from diffusers.pipelines.controlnet.pipeline_controlnet.StableDiffusionControlNetPipeline.check_image
+ def check_image(self, image, prompt, prompt_embeds):
+ image_is_pil = isinstance(image, PIL.Image.Image)
+ image_is_tensor = isinstance(image, torch.Tensor)
+ image_is_np = isinstance(image, np.ndarray)
+ image_is_pil_list = isinstance(image, list) and isinstance(image[0], PIL.Image.Image)
+ image_is_tensor_list = isinstance(image, list) and isinstance(image[0], torch.Tensor)
+ image_is_np_list = isinstance(image, list) and isinstance(image[0], np.ndarray)
+
+ if (
+ not image_is_pil
+ and not image_is_tensor
+ and not image_is_np
+ and not image_is_pil_list
+ and not image_is_tensor_list
+ and not image_is_np_list
+ ):
+ raise TypeError(
+ f"image must be passed and be one of PIL image, numpy array, torch tensor, list of PIL images, list of numpy arrays or list of torch tensors, but is {type(image)}"
+ )
+
+ if image_is_pil:
+ image_batch_size = 1
+ else:
+ image_batch_size = len(image)
+
+ if prompt is not None and isinstance(prompt, str):
+ prompt_batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ prompt_batch_size = len(prompt)
+ elif prompt_embeds is not None:
+ prompt_batch_size = prompt_embeds.shape[0]
+
+ if image_batch_size != 1 and image_batch_size != prompt_batch_size:
+ raise ValueError(
+ f"If image batch size is not 1, image batch size must be same as prompt batch size. image batch size: {image_batch_size}, prompt batch size: {prompt_batch_size}"
+ )
+
+ # Copied from diffusers.pipelines.text_to_video_synthesis.pipeline_text_to_video_synth.TextToVideoSDPipeline.prepare_latents
+ def prepare_latents(
+ self, batch_size, num_channels_latents, num_frames, height, width, dtype, device, generator, latents=None
+ ):
+ shape = (
+ batch_size,
+ num_channels_latents,
+ num_frames,
+ height // self.vae_scale_factor,
+ width // self.vae_scale_factor,
+ )
+ if isinstance(generator, list) and len(generator) != batch_size:
+ raise ValueError(
+ f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+ f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+ )
+
+ if latents is None:
+ latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+ else:
+ latents = latents.to(device)
+
+ # scale the initial noise by the standard deviation required by the scheduler
+ latents = latents * self.scheduler.init_noise_sigma
+ return latents
+
+ def prepare_image(self, image, width, height, device, dtype):
+ image = self.control_image_processor.preprocess(image, height=height, width=width)
+ controlnet_images = image.unsqueeze(0).to(device, dtype)
+ batch_size, num_frames, channels, height, width = controlnet_images.shape
+
+ # TODO: remove below line
+ assert controlnet_images.min() >= 0 and controlnet_images.max() <= 1
+
+ if self.controlnet.use_simplified_condition_embedding:
+ controlnet_images = controlnet_images.reshape(batch_size * num_frames, channels, height, width)
+ controlnet_images = 2 * controlnet_images - 1
+ conditioning_frames = retrieve_latents(self.vae.encode(controlnet_images)) * self.vae.config.scaling_factor
+ conditioning_frames = conditioning_frames.reshape(
+ batch_size, num_frames, 4, height // self.vae_scale_factor, width // self.vae_scale_factor
+ )
+ else:
+ conditioning_frames = controlnet_images
+
+ conditioning_frames = conditioning_frames.permute(0, 2, 1, 3, 4) # [b, c, f, h, w]
+ return conditioning_frames
+
+ def prepare_sparse_control_conditioning(
+ self,
+ conditioning_frames: torch.Tensor,
+ num_frames: int,
+ controlnet_frame_indices: int,
+ device: torch.device,
+ dtype: torch.dtype,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ assert conditioning_frames.shape[2] >= len(controlnet_frame_indices)
+
+ batch_size, channels, _, height, width = conditioning_frames.shape
+ controlnet_cond = torch.zeros((batch_size, channels, num_frames, height, width), dtype=dtype, device=device)
+ controlnet_cond_mask = torch.zeros((batch_size, 1, num_frames, height, width), dtype=dtype, device=device)
+ controlnet_cond[:, :, controlnet_frame_indices] = conditioning_frames[:, :, : len(controlnet_frame_indices)]
+ controlnet_cond_mask[:, :, controlnet_frame_indices] = 1
+
+ return controlnet_cond, controlnet_cond_mask
+
+ @property
+ def guidance_scale(self):
+ return self._guidance_scale
+
+ @property
+ def clip_skip(self):
+ return self._clip_skip
+
+ # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+ # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+ # corresponds to doing no classifier free guidance.
+ @property
+ def do_classifier_free_guidance(self):
+ return self._guidance_scale > 1
+
+ @property
+ def cross_attention_kwargs(self):
+ return self._cross_attention_kwargs
+
+ @property
+ def num_timesteps(self):
+ return self._num_timesteps
+
+ @torch.no_grad()
+ @replace_example_docstring(EXAMPLE_DOC_STRING)
+ def __call__(
+ self,
+ prompt: Optional[Union[str, List[str]]] = None,
+ height: Optional[int] = None,
+ width: Optional[int] = None,
+ num_frames: int = 16,
+ num_inference_steps: int = 50,
+ guidance_scale: float = 7.5,
+ negative_prompt: Optional[Union[str, List[str]]] = None,
+ num_videos_per_prompt: int = 1,
+ eta: float = 0.0,
+ generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+ latents: Optional[torch.Tensor] = None,
+ prompt_embeds: Optional[torch.Tensor] = None,
+ negative_prompt_embeds: Optional[torch.Tensor] = None,
+ ip_adapter_image: Optional[PipelineImageInput] = None,
+ ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
+ conditioning_frames: Optional[List[PipelineImageInput]] = None,
+ output_type: str = "pil",
+ return_dict: bool = True,
+ cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+ controlnet_conditioning_scale: Union[float, List[float]] = 1.0,
+ controlnet_frame_indices: List[int] = [0],
+ guess_mode: bool = False,
+ clip_skip: Optional[int] = None,
+ callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+ callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+ ):
+ r"""
+ The call function to the pipeline for generation.
+
+ Args:
+ prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+ height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The height in pixels of the generated video.
+ width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`):
+ The width in pixels of the generated video.
+ num_frames (`int`, *optional*, defaults to 16):
+ The number of video frames that are generated. Defaults to 16 frames which at 8 frames per seconds
+ amounts to 2 seconds of video.
+ num_inference_steps (`int`, *optional*, defaults to 50):
+ The number of denoising steps. More denoising steps usually lead to a higher quality videos at the
+ expense of slower inference.
+ guidance_scale (`float`, *optional*, defaults to 7.5):
+ A higher guidance scale value encourages the model to generate images closely linked to the text
+ `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+ negative_prompt (`str` or `List[str]`, *optional*):
+ The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+ pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`).
+ eta (`float`, *optional*, defaults to 0.0):
+ Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+ to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+ generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+ A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
+ generation deterministic.
+ latents (`torch.Tensor`, *optional*):
+ Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for video
+ generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+ tensor is generated by sampling using the supplied random `generator`. Latents should be of shape
+ `(batch_size, num_channel, num_frames, height, width)`.
+ prompt_embeds (`torch.Tensor`, *optional*):
+ Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not
+ provided, text embeddings are generated from the `prompt` input argument.
+ negative_prompt_embeds (`torch.Tensor`, *optional*):
+ Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+ not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+ ip_adapter_image: (`PipelineImageInput`, *optional*):
+ Optional image input to work with IP Adapters.
+ ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
+ Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
+ IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should
+ contain the negative image embedding if `do_classifier_free_guidance` is set to `True`. If not
+ provided, embeddings are computed from the `ip_adapter_image` input argument.
+ conditioning_frames (`List[PipelineImageInput]`, *optional*):
+ The SparseControlNet input to provide guidance to the `unet` for generation.
+ output_type (`str`, *optional*, defaults to `"pil"`):
+ The output format of the generated video. Choose between `torch.Tensor`, `PIL.Image` or `np.array`.
+ return_dict (`bool`, *optional*, defaults to `True`):
+ Whether or not to return a [`~pipelines.text_to_video_synthesis.TextToVideoSDPipelineOutput`] instead
+ of a plain tuple.
+ cross_attention_kwargs (`dict`, *optional*):
+ A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+ [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+ controlnet_conditioning_scale (`float` or `List[float]`, *optional*, defaults to 1.0):
+ The outputs of the ControlNet are multiplied by `controlnet_conditioning_scale` before they are added
+ to the residual in the original `unet`. If multiple ControlNets are specified in `init`, you can set
+ the corresponding scale as a list.
+ controlnet_frame_indices (`List[int]`):
+ The indices where the conditioning frames must be applied for generation. Multiple frames can be
+ provided to guide the model to generate similar structure outputs, where the `unet` can
+ "fill-in-the-gaps" for interpolation videos, or a single frame could be provided for general expected
+ structure. Must have the same length as `conditioning_frames`.
+ clip_skip (`int`, *optional*):
+ Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+ the output of the pre-final layer will be used for computing the prompt embeddings.
+ callback_on_step_end (`Callable`, *optional*):
+ A function that calls at the end of each denoising steps during the inference. The function is called
+ with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+ callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+ `callback_on_step_end_tensor_inputs`.
+ callback_on_step_end_tensor_inputs (`List`, *optional*):
+ The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+ will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+ `._callback_tensor_inputs` attribute of your pipeline class.
+
+ Examples:
+
+ Returns:
+ [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] or `tuple`:
+ If `return_dict` is `True`, [`~pipelines.animatediff.pipeline_output.AnimateDiffPipelineOutput`] is
+ returned, otherwise a `tuple` is returned where the first element is a list with the generated frames.
+ """
+ controlnet = self.controlnet._orig_mod if is_compiled_module(self.controlnet) else self.controlnet
+
+ # 0. Default height and width to unet
+ height = height or self.unet.config.sample_size * self.vae_scale_factor
+ width = width or self.unet.config.sample_size * self.vae_scale_factor
+ num_videos_per_prompt = 1
+
+ # 1. Check inputs. Raise error if not correct
+ self.check_inputs(
+ prompt=prompt,
+ height=height,
+ width=width,
+ negative_prompt=negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ ip_adapter_image=ip_adapter_image,
+ ip_adapter_image_embeds=ip_adapter_image_embeds,
+ callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+ image=conditioning_frames,
+ controlnet_conditioning_scale=controlnet_conditioning_scale,
+ )
+
+ self._guidance_scale = guidance_scale
+ self._clip_skip = clip_skip
+ self._cross_attention_kwargs = cross_attention_kwargs
+
+ # 2. Define call parameters
+ if prompt is not None and isinstance(prompt, str):
+ batch_size = 1
+ elif prompt is not None and isinstance(prompt, list):
+ batch_size = len(prompt)
+ else:
+ batch_size = prompt_embeds.shape[0]
+
+ device = self._execution_device
+
+ global_pool_conditions = (
+ controlnet.config.global_pool_conditions
+ if isinstance(controlnet, SparseControlNetModel)
+ else controlnet.nets[0].config.global_pool_conditions
+ )
+ guess_mode = guess_mode or global_pool_conditions
+
+ # 3. Encode input prompt
+ text_encoder_lora_scale = (
+ self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+ )
+ prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+ prompt,
+ device,
+ num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ negative_prompt,
+ prompt_embeds=prompt_embeds,
+ negative_prompt_embeds=negative_prompt_embeds,
+ lora_scale=text_encoder_lora_scale,
+ clip_skip=self.clip_skip,
+ )
+ # For classifier free guidance, we need to do two forward passes.
+ # Here we concatenate the unconditional and text embeddings into a single batch
+ # to avoid doing two forward passes
+ if self.do_classifier_free_guidance:
+ prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+ # 4. Prepare IP-Adapter embeddings
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+ image_embeds = self.prepare_ip_adapter_image_embeds(
+ ip_adapter_image,
+ ip_adapter_image_embeds,
+ device,
+ batch_size * num_videos_per_prompt,
+ self.do_classifier_free_guidance,
+ )
+
+ # 5. Prepare controlnet conditioning
+ conditioning_frames = self.prepare_image(conditioning_frames, width, height, device, controlnet.dtype)
+ controlnet_cond, controlnet_cond_mask = self.prepare_sparse_control_conditioning(
+ conditioning_frames, num_frames, controlnet_frame_indices, device, controlnet.dtype
+ )
+
+ # 6. Prepare timesteps
+ self.scheduler.set_timesteps(num_inference_steps, device=device)
+ timesteps = self.scheduler.timesteps
+
+ # 7. Prepare latent variables
+ num_channels_latents = self.unet.config.in_channels
+ latents = self.prepare_latents(
+ batch_size * num_videos_per_prompt,
+ num_channels_latents,
+ num_frames,
+ height,
+ width,
+ prompt_embeds.dtype,
+ device,
+ generator,
+ latents,
+ )
+
+ # 8. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+ extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+ # 9. Add image embeds for IP-Adapter
+ added_cond_kwargs = (
+ {"image_embeds": image_embeds}
+ if ip_adapter_image is not None or ip_adapter_image_embeds is not None
+ else None
+ )
+
+ num_free_init_iters = self._free_init_num_iters if self.free_init_enabled else 1
+ for free_init_iter in range(num_free_init_iters):
+ if self.free_init_enabled:
+ latents, timesteps = self._apply_free_init(
+ latents, free_init_iter, num_inference_steps, device, latents.dtype, generator
+ )
+
+ self._num_timesteps = len(timesteps)
+ num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+
+ # 10. Denoising loop
+ with self.progress_bar(total=self._num_timesteps) as progress_bar:
+ for i, t in enumerate(timesteps):
+ # expand the latents if we are doing classifier free guidance
+ latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+ latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+ if guess_mode and self.do_classifier_free_guidance:
+ # Infer SparseControlNetModel only for the conditional batch.
+ control_model_input = latents
+ control_model_input = self.scheduler.scale_model_input(control_model_input, t)
+ controlnet_prompt_embeds = prompt_embeds.chunk(2)[1]
+ else:
+ control_model_input = latent_model_input
+ controlnet_prompt_embeds = prompt_embeds
+
+ down_block_res_samples, mid_block_res_sample = self.controlnet(
+ control_model_input,
+ t,
+ encoder_hidden_states=controlnet_prompt_embeds,
+ controlnet_cond=controlnet_cond,
+ conditioning_mask=controlnet_cond_mask,
+ conditioning_scale=controlnet_conditioning_scale,
+ guess_mode=guess_mode,
+ return_dict=False,
+ )
+
+ # predict the noise residual
+ noise_pred = self.unet(
+ latent_model_input,
+ t,
+ encoder_hidden_states=prompt_embeds,
+ cross_attention_kwargs=cross_attention_kwargs,
+ added_cond_kwargs=added_cond_kwargs,
+ down_block_additional_residuals=down_block_res_samples,
+ mid_block_additional_residual=mid_block_res_sample,
+ ).sample
+
+ # perform guidance
+ if self.do_classifier_free_guidance:
+ noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+ noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+ # compute the previous noisy sample x_t -> x_t-1
+ latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs).prev_sample
+
+ if callback_on_step_end is not None:
+ callback_kwargs = {}
+ for k in callback_on_step_end_tensor_inputs:
+ callback_kwargs[k] = locals()[k]
+ callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+ latents = callback_outputs.pop("latents", latents)
+ prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+ negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+ # call the callback, if provided
+ if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+ progress_bar.update()
+
+ # 11. Post processing
+ if output_type == "latent":
+ video = latents
+ else:
+ video_tensor = self.decode_latents(latents)
+ video = self.video_processor.postprocess_video(video=video_tensor, output_type=output_type)
+
+ # 12. Offload all models
+ self.maybe_free_model_hooks()
+
+ if not return_dict:
+ return (video,)
+
+ return AnimateDiffPipelineOutput(frames=video)
diff --git a/src/diffusers/utils/dummy_pt_objects.py b/src/diffusers/utils/dummy_pt_objects.py
index 5df0d6d28f53..3ead6fd99d10 100644
--- a/src/diffusers/utils/dummy_pt_objects.py
+++ b/src/diffusers/utils/dummy_pt_objects.py
@@ -362,6 +362,21 @@ def from_pretrained(cls, *args, **kwargs):
requires_backends(cls, ["torch"])
+class SparseControlNetModel(metaclass=DummyObject):
+ _backends = ["torch"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch"])
+
+
class T2IAdapter(metaclass=DummyObject):
_backends = ["torch"]
diff --git a/src/diffusers/utils/dummy_torch_and_transformers_objects.py b/src/diffusers/utils/dummy_torch_and_transformers_objects.py
index 399656d8c185..8e40e5128854 100644
--- a/src/diffusers/utils/dummy_torch_and_transformers_objects.py
+++ b/src/diffusers/utils/dummy_torch_and_transformers_objects.py
@@ -107,6 +107,21 @@ def from_pretrained(cls, *args, **kwargs):
requires_backends(cls, ["torch", "transformers"])
+class AnimateDiffSparseControlNetPipeline(metaclass=DummyObject):
+ _backends = ["torch", "transformers"]
+
+ def __init__(self, *args, **kwargs):
+ requires_backends(self, ["torch", "transformers"])
+
+ @classmethod
+ def from_config(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+ @classmethod
+ def from_pretrained(cls, *args, **kwargs):
+ requires_backends(cls, ["torch", "transformers"])
+
+
class AnimateDiffVideoToVideoPipeline(metaclass=DummyObject):
_backends = ["torch", "transformers"]
diff --git a/tests/pipelines/animatediff/test_animatediff_sparsectrl.py b/tests/pipelines/animatediff/test_animatediff_sparsectrl.py
new file mode 100644
index 000000000000..5d8a7228118d
--- /dev/null
+++ b/tests/pipelines/animatediff/test_animatediff_sparsectrl.py
@@ -0,0 +1,478 @@
+import unittest
+
+import numpy as np
+import torch
+from PIL import Image
+from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
+
+import diffusers
+from diffusers import (
+ AnimateDiffSparseControlNetPipeline,
+ AutoencoderKL,
+ DDIMScheduler,
+ DPMSolverMultistepScheduler,
+ LCMScheduler,
+ MotionAdapter,
+ SparseControlNetModel,
+ StableDiffusionPipeline,
+ UNet2DConditionModel,
+ UNetMotionModel,
+)
+from diffusers.utils import logging
+from diffusers.utils.testing_utils import torch_device
+
+from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_PARAMS
+from ..test_pipelines_common import (
+ IPAdapterTesterMixin,
+ PipelineFromPipeTesterMixin,
+ PipelineTesterMixin,
+ SDFunctionTesterMixin,
+)
+
+
+def to_np(tensor):
+ if isinstance(tensor, torch.Tensor):
+ tensor = tensor.detach().cpu().numpy()
+
+ return tensor
+
+
+class AnimateDiffSparseControlNetPipelineFastTests(
+ IPAdapterTesterMixin, SDFunctionTesterMixin, PipelineTesterMixin, PipelineFromPipeTesterMixin, unittest.TestCase
+):
+ pipeline_class = AnimateDiffSparseControlNetPipeline
+ params = TEXT_TO_IMAGE_PARAMS
+ batch_params = TEXT_TO_IMAGE_BATCH_PARAMS
+ required_optional_params = frozenset(
+ [
+ "num_inference_steps",
+ "generator",
+ "latents",
+ "return_dict",
+ "callback_on_step_end",
+ "callback_on_step_end_tensor_inputs",
+ ]
+ )
+
+ def get_dummy_components(self):
+ cross_attention_dim = 8
+ block_out_channels = (8, 8)
+
+ torch.manual_seed(0)
+ unet = UNet2DConditionModel(
+ block_out_channels=block_out_channels,
+ layers_per_block=2,
+ sample_size=8,
+ in_channels=4,
+ out_channels=4,
+ down_block_types=("CrossAttnDownBlock2D", "DownBlock2D"),
+ up_block_types=("CrossAttnUpBlock2D", "UpBlock2D"),
+ cross_attention_dim=cross_attention_dim,
+ norm_num_groups=2,
+ )
+ scheduler = DDIMScheduler(
+ beta_start=0.00085,
+ beta_end=0.012,
+ beta_schedule="linear",
+ clip_sample=False,
+ )
+ torch.manual_seed(0)
+ controlnet = SparseControlNetModel(
+ block_out_channels=block_out_channels,
+ layers_per_block=2,
+ in_channels=4,
+ conditioning_channels=3,
+ down_block_types=("CrossAttnDownBlockMotion", "DownBlockMotion"),
+ cross_attention_dim=cross_attention_dim,
+ conditioning_embedding_out_channels=(8, 8),
+ norm_num_groups=1,
+ use_simplified_condition_embedding=False,
+ )
+ torch.manual_seed(0)
+ vae = AutoencoderKL(
+ block_out_channels=block_out_channels,
+ in_channels=3,
+ out_channels=3,
+ down_block_types=["DownEncoderBlock2D", "DownEncoderBlock2D"],
+ up_block_types=["UpDecoderBlock2D", "UpDecoderBlock2D"],
+ latent_channels=4,
+ norm_num_groups=2,
+ )
+ torch.manual_seed(0)
+ text_encoder_config = CLIPTextConfig(
+ bos_token_id=0,
+ eos_token_id=2,
+ hidden_size=cross_attention_dim,
+ intermediate_size=37,
+ layer_norm_eps=1e-05,
+ num_attention_heads=4,
+ num_hidden_layers=5,
+ pad_token_id=1,
+ vocab_size=1000,
+ )
+ text_encoder = CLIPTextModel(text_encoder_config)
+ tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+ motion_adapter = MotionAdapter(
+ block_out_channels=block_out_channels,
+ motion_layers_per_block=2,
+ motion_norm_num_groups=2,
+ motion_num_attention_heads=4,
+ )
+
+ components = {
+ "unet": unet,
+ "controlnet": controlnet,
+ "scheduler": scheduler,
+ "vae": vae,
+ "motion_adapter": motion_adapter,
+ "text_encoder": text_encoder,
+ "tokenizer": tokenizer,
+ "feature_extractor": None,
+ "image_encoder": None,
+ }
+ return components
+
+ def get_dummy_inputs(self, device, seed: int = 0, num_frames: int = 2):
+ if str(device).startswith("mps"):
+ generator = torch.manual_seed(seed)
+ else:
+ generator = torch.Generator(device=device).manual_seed(seed)
+
+ video_height = 32
+ video_width = 32
+ conditioning_frames = [Image.new("RGB", (video_width, video_height))] * num_frames
+
+ inputs = {
+ "prompt": "A painting of a squirrel eating a burger",
+ "conditioning_frames": conditioning_frames,
+ "controlnet_frame_indices": list(range(num_frames)),
+ "generator": generator,
+ "num_inference_steps": 2,
+ "num_frames": num_frames,
+ "guidance_scale": 7.5,
+ "output_type": "pt",
+ }
+ return inputs
+
+ def test_from_pipe_consistent_config(self):
+ assert self.original_pipeline_class == StableDiffusionPipeline
+ original_repo = "hf-internal-testing/tinier-stable-diffusion-pipe"
+ original_kwargs = {"requires_safety_checker": False}
+
+ # create original_pipeline_class(sd)
+ pipe_original = self.original_pipeline_class.from_pretrained(original_repo, **original_kwargs)
+
+ # original_pipeline_class(sd) -> pipeline_class
+ pipe_components = self.get_dummy_components()
+ pipe_additional_components = {}
+ for name, component in pipe_components.items():
+ if name not in pipe_original.components:
+ pipe_additional_components[name] = component
+
+ pipe = self.pipeline_class.from_pipe(pipe_original, **pipe_additional_components)
+
+ # pipeline_class -> original_pipeline_class(sd)
+ original_pipe_additional_components = {}
+ for name, component in pipe_original.components.items():
+ if name not in pipe.components or not isinstance(component, pipe.components[name].__class__):
+ original_pipe_additional_components[name] = component
+
+ pipe_original_2 = self.original_pipeline_class.from_pipe(pipe, **original_pipe_additional_components)
+
+ # compare the config
+ original_config = {k: v for k, v in pipe_original.config.items() if not k.startswith("_")}
+ original_config_2 = {k: v for k, v in pipe_original_2.config.items() if not k.startswith("_")}
+ assert original_config_2 == original_config
+
+ def test_motion_unet_loading(self):
+ components = self.get_dummy_components()
+ pipe = AnimateDiffSparseControlNetPipeline(**components)
+
+ assert isinstance(pipe.unet, UNetMotionModel)
+
+ @unittest.skip("Attention slicing is not enabled in this pipeline")
+ def test_attention_slicing_forward_pass(self):
+ pass
+
+ def test_ip_adapter_single(self):
+ expected_pipe_slice = None
+ if torch_device == "cpu":
+ expected_pipe_slice = np.array(
+ [
+ 0.6604,
+ 0.4099,
+ 0.4928,
+ 0.5706,
+ 0.5096,
+ 0.5012,
+ 0.6051,
+ 0.5169,
+ 0.5021,
+ 0.4864,
+ 0.4261,
+ 0.5779,
+ 0.5822,
+ 0.4049,
+ 0.5253,
+ 0.6160,
+ 0.4150,
+ 0.5155,
+ ]
+ )
+ return super().test_ip_adapter_single(expected_pipe_slice=expected_pipe_slice)
+
+ def test_dict_tuple_outputs_equivalent(self):
+ expected_slice = None
+ if torch_device == "cpu":
+ expected_slice = np.array([0.6051, 0.5169, 0.5021, 0.6160, 0.4150, 0.5155])
+ return super().test_dict_tuple_outputs_equivalent(expected_slice=expected_slice)
+
+ def test_inference_batch_single_identical(
+ self,
+ batch_size=2,
+ expected_max_diff=1e-4,
+ additional_params_copy_to_batched_inputs=["num_inference_steps"],
+ ):
+ components = self.get_dummy_components()
+ pipe = self.pipeline_class(**components)
+ for components in pipe.components.values():
+ if hasattr(components, "set_default_attn_processor"):
+ components.set_default_attn_processor()
+
+ pipe.to(torch_device)
+ pipe.set_progress_bar_config(disable=None)
+ inputs = self.get_dummy_inputs(torch_device)
+ # Reset generator in case it is has been used in self.get_dummy_inputs
+ inputs["generator"] = self.get_generator(0)
+
+ logger = logging.get_logger(pipe.__module__)
+ logger.setLevel(level=diffusers.logging.FATAL)
+
+ # batchify inputs
+ batched_inputs = {}
+ batched_inputs.update(inputs)
+
+ for name in self.batch_params:
+ if name not in inputs:
+ continue
+
+ value = inputs[name]
+ if name == "prompt":
+ len_prompt = len(value)
+ batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
+ batched_inputs[name][-1] = 100 * "very long"
+
+ else:
+ batched_inputs[name] = batch_size * [value]
+
+ if "generator" in inputs:
+ batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]
+
+ if "batch_size" in inputs:
+ batched_inputs["batch_size"] = batch_size
+
+ for arg in additional_params_copy_to_batched_inputs:
+ batched_inputs[arg] = inputs[arg]
+
+ output = pipe(**inputs)
+ output_batch = pipe(**batched_inputs)
+
+ assert output_batch[0].shape[0] == batch_size
+
+ max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
+ assert max_diff < expected_max_diff
+
+ def test_inference_batch_single_identical_use_simplified_condition_embedding_true(
+ self,
+ batch_size=2,
+ expected_max_diff=1e-4,
+ additional_params_copy_to_batched_inputs=["num_inference_steps"],
+ ):
+ components = self.get_dummy_components()
+
+ torch.manual_seed(0)
+ old_controlnet = components.pop("controlnet")
+ components["controlnet"] = SparseControlNetModel.from_config(
+ old_controlnet.config, conditioning_channels=4, use_simplified_condition_embedding=True
+ )
+
+ pipe = self.pipeline_class(**components)
+ for components in pipe.components.values():
+ if hasattr(components, "set_default_attn_processor"):
+ components.set_default_attn_processor()
+
+ pipe.to(torch_device)
+ pipe.set_progress_bar_config(disable=None)
+ inputs = self.get_dummy_inputs(torch_device)
+ # Reset generator in case it is has been used in self.get_dummy_inputs
+ inputs["generator"] = self.get_generator(0)
+
+ logger = logging.get_logger(pipe.__module__)
+ logger.setLevel(level=diffusers.logging.FATAL)
+
+ # batchify inputs
+ batched_inputs = {}
+ batched_inputs.update(inputs)
+
+ for name in self.batch_params:
+ if name not in inputs:
+ continue
+
+ value = inputs[name]
+ if name == "prompt":
+ len_prompt = len(value)
+ batched_inputs[name] = [value[: len_prompt // i] for i in range(1, batch_size + 1)]
+ batched_inputs[name][-1] = 100 * "very long"
+
+ else:
+ batched_inputs[name] = batch_size * [value]
+
+ if "generator" in inputs:
+ batched_inputs["generator"] = [self.get_generator(i) for i in range(batch_size)]
+
+ if "batch_size" in inputs:
+ batched_inputs["batch_size"] = batch_size
+
+ for arg in additional_params_copy_to_batched_inputs:
+ batched_inputs[arg] = inputs[arg]
+
+ output = pipe(**inputs)
+ output_batch = pipe(**batched_inputs)
+
+ assert output_batch[0].shape[0] == batch_size
+
+ max_diff = np.abs(to_np(output_batch[0][0]) - to_np(output[0][0])).max()
+ assert max_diff < expected_max_diff
+
+ @unittest.skipIf(torch_device != "cuda", reason="CUDA and CPU are required to switch devices")
+ def test_to_device(self):
+ components = self.get_dummy_components()
+ pipe = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+
+ pipe.to("cpu")
+ # pipeline creates a new motion UNet under the hood. So we need to check the device from pipe.components
+ model_devices = [
+ component.device.type for component in pipe.components.values() if hasattr(component, "device")
+ ]
+ self.assertTrue(all(device == "cpu" for device in model_devices))
+
+ output_cpu = pipe(**self.get_dummy_inputs("cpu"))[0]
+ self.assertTrue(np.isnan(output_cpu).sum() == 0)
+
+ pipe.to("cuda")
+ model_devices = [
+ component.device.type for component in pipe.components.values() if hasattr(component, "device")
+ ]
+ self.assertTrue(all(device == "cuda" for device in model_devices))
+
+ output_cuda = pipe(**self.get_dummy_inputs("cuda"))[0]
+ self.assertTrue(np.isnan(to_np(output_cuda)).sum() == 0)
+
+ def test_to_dtype(self):
+ components = self.get_dummy_components()
+ pipe = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+
+ # pipeline creates a new motion UNet under the hood. So we need to check the dtype from pipe.components
+ model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
+ self.assertTrue(all(dtype == torch.float32 for dtype in model_dtypes))
+
+ pipe.to(dtype=torch.float16)
+ model_dtypes = [component.dtype for component in pipe.components.values() if hasattr(component, "dtype")]
+ self.assertTrue(all(dtype == torch.float16 for dtype in model_dtypes))
+
+ def test_prompt_embeds(self):
+ components = self.get_dummy_components()
+ pipe = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+ pipe.to(torch_device)
+
+ inputs = self.get_dummy_inputs(torch_device)
+ inputs.pop("prompt")
+ inputs["prompt_embeds"] = torch.randn((1, 4, pipe.text_encoder.config.hidden_size), device=torch_device)
+ pipe(**inputs)
+
+ def test_free_init(self):
+ components = self.get_dummy_components()
+ pipe: AnimateDiffSparseControlNetPipeline = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+ pipe.to(torch_device)
+
+ inputs_normal = self.get_dummy_inputs(torch_device)
+ frames_normal = pipe(**inputs_normal).frames[0]
+
+ pipe.enable_free_init(
+ num_iters=2,
+ use_fast_sampling=True,
+ method="butterworth",
+ order=4,
+ spatial_stop_frequency=0.25,
+ temporal_stop_frequency=0.25,
+ )
+ inputs_enable_free_init = self.get_dummy_inputs(torch_device)
+ frames_enable_free_init = pipe(**inputs_enable_free_init).frames[0]
+
+ pipe.disable_free_init()
+ inputs_disable_free_init = self.get_dummy_inputs(torch_device)
+ frames_disable_free_init = pipe(**inputs_disable_free_init).frames[0]
+
+ sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_init)).sum()
+ max_diff_disabled = np.abs(to_np(frames_normal) - to_np(frames_disable_free_init)).max()
+ self.assertGreater(
+ sum_enabled, 1e1, "Enabling of FreeInit should lead to results different from the default pipeline results"
+ )
+ self.assertLess(
+ max_diff_disabled,
+ 1e-4,
+ "Disabling of FreeInit should lead to results similar to the default pipeline results",
+ )
+
+ def test_free_init_with_schedulers(self):
+ components = self.get_dummy_components()
+ pipe: AnimateDiffSparseControlNetPipeline = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+ pipe.to(torch_device)
+
+ inputs_normal = self.get_dummy_inputs(torch_device)
+ frames_normal = pipe(**inputs_normal).frames[0]
+
+ schedulers_to_test = [
+ DPMSolverMultistepScheduler.from_config(
+ components["scheduler"].config,
+ timestep_spacing="linspace",
+ beta_schedule="linear",
+ algorithm_type="dpmsolver++",
+ steps_offset=1,
+ clip_sample=False,
+ ),
+ LCMScheduler.from_config(
+ components["scheduler"].config,
+ timestep_spacing="linspace",
+ beta_schedule="linear",
+ steps_offset=1,
+ clip_sample=False,
+ ),
+ ]
+ components.pop("scheduler")
+
+ for scheduler in schedulers_to_test:
+ components["scheduler"] = scheduler
+ pipe: AnimateDiffSparseControlNetPipeline = self.pipeline_class(**components)
+ pipe.set_progress_bar_config(disable=None)
+ pipe.to(torch_device)
+
+ pipe.enable_free_init(num_iters=2, use_fast_sampling=False)
+
+ inputs = self.get_dummy_inputs(torch_device)
+ frames_enable_free_init = pipe(**inputs).frames[0]
+ sum_enabled = np.abs(to_np(frames_normal) - to_np(frames_enable_free_init)).sum()
+
+ self.assertGreater(
+ sum_enabled,
+ 1e1,
+ "Enabling of FreeInit should lead to results different from the default pipeline results",
+ )
+
+ def test_vae_slicing(self):
+ return super().test_vae_slicing(image_count=2)