LLM/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2025-12-06 12:34:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

[core] AnimateDiff SparseCtrl (#8897)

Browse Source 
* initial sparse control model draft

* remove unnecessary implementation

* copy animatediff pipeline

* remove deprecated callbacks

* update

* update pipeline implementation progress

* make style

* make fix-copies

* update progress

* add partially working pipeline

* remove debug prints

* add model docs

* dummy objects

* improve motion lora conversion script

* fix bugs

* update docstrings

* remove unnecessary model params; docs

* address review comment

* add copied from to zero_module

* copy animatediff test

* add fast tests

* update docs

* update

* update pipeline docs

* fix expected slice values

* fix license

* remove get_down_block usage

* remove temporal_double_self_attention from get_down_block

* update

* update docs with org and documentation images

* make from_unet work in sparsecontrolnetmodel

* add latest freeinit test from #8969

* make fix-copies

* LoraLoaderMixin -> StableDiffsuionLoraLoaderMixin
This commit is contained in:
Aryan
2024-07-26 17:46:05 +05:30
committed by GitHub
parent 57a021d5e4
commit 5c53ca5ed8
16 changed files with 2664 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/en/_toctree.yml
View File

@@ -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:

46
docs/source/en/api/models/controlnet_sparsectrl.md Normal file
View File

@@ -0,0 +1,46 @@
<!-- 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. -->
# 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

190
docs/source/en/api/pipelines/animatediff.md
View File

@@ -100,6 +100,189 @@ AnimateDiff tends to work better with finetuned Stable Diffusion models. If you
</Tip>
### 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:
<table align="center">
<tr>
<center>
<b>an aerial view of a cyberpunk city, night time, neon lights, masterpiece, high quality</b>
</center>
</tr>
<tr>
<td>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-1.png" alt="scribble-1" />
</center>
</td>
<td>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-2.png" alt="scribble-2" />
</center>
</td>
<td>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-scribble-3.png" alt="scribble-3" />
</center>
</td>
</tr>
<tr>
<td colspan=3>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-sparsectrl-scribble-results.gif" alt="an aerial view of a cyberpunk city, night time, neon lights, masterpiece, high quality" />
</center>
</td>
</tr>
</table>
#### 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:
<table align="center">
<tr>
<center>
<b>closeup face photo of man in black clothes, night city street, bokeh, fireworks in background</b>
</center>
</tr>
<tr>
<td>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-firework.png" alt="closeup face photo of man in black clothes, night city street, bokeh, fireworks in background" />
</center>
</td>
<td>
<center>
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/animatediff-sparsectrl-rgb-result.gif" alt="closeup face photo of man in black clothes, night city street, bokeh, fireworks in background" />
</center>
</td>
</tr>
</table>
### 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

21
scripts/convert_animatediff_motion_lora_to_diffusers.py
View File

@@ -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 @@ if __name__ == "__main__":
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")

83
scripts/convert_animatediff_sparsectrl_to_diffusers.py Normal file
View File

@@ -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)

4
src/diffusers/__init__.py
View File

@@ -99,6 +99,7 @@ else:
"SD3ControlNetModel",
"SD3MultiControlNetModel",
"SD3Transformer2DModel",
"SparseControlNetModel",
"StableCascadeUNet",
"T2IAdapter",
"T5FilmDecoder",
@@ -231,6 +232,7 @@ else:
"AmusedPipeline",
"AnimateDiffPipeline",
"AnimateDiffSDXLPipeline",
"AnimateDiffSparseControlNetPipeline",
"AnimateDiffVideoToVideoPipeline",
"AudioLDM2Pipeline",
"AudioLDM2ProjectionModel",
@@ -533,6 +535,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
SD3ControlNetModel,
SD3MultiControlNetModel,
SD3Transformer2DModel,
SparseControlNetModel,
T2IAdapter,
T5FilmDecoder,
Transformer2DModel,
@@ -645,6 +648,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
AmusedPipeline,
AnimateDiffPipeline,
AnimateDiffSDXLPipeline,
AnimateDiffSparseControlNetPipeline,
AnimateDiffVideoToVideoPipeline,
AudioLDM2Pipeline,
AudioLDM2ProjectionModel,

2
src/diffusers/models/__init__.py
View File

@@ -35,6 +35,7 @@ if is_torch_available():
_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 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
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

1
src/diffusers/models/controlnet.py
View File

@@ -830,7 +830,6 @@ class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalModelMixin):
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):

791
src/diffusers/models/controlnet_sparsectrl.py Normal file
View File

@@ -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

4
src/diffusers/models/unets/unet_3d_blocks.py
View File

@@ -966,6 +966,7 @@ class DownBlockMotion(nn.Module):
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 @@ class DownBlockMotion(nn.Module):
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 @@ class CrossAttnDownBlockMotion(nn.Module):
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 @@ class CrossAttnDownBlockMotion(nn.Module):
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,
)
)

8
src/diffusers/pipelines/__init__.py
View File

@@ -119,6 +119,7 @@ else:
_import_structure["animatediff"] = [
"AnimateDiffPipeline",
"AnimateDiffSDXLPipeline",
"AnimateDiffSparseControlNetPipeline",
"AnimateDiffVideoToVideoPipeline",
]
_import_structure["audioldm"] = ["AudioLDMPipeline"]
@@ -413,7 +414,12 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
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,

2
src/diffusers/pipelines/animatediff/__init__.py
View File

@@ -23,6 +23,7 @@ except OptionalDependencyNotAvailable:
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 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
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

1008
src/diffusers/pipelines/animatediff/pipeline_animatediff_sparsectrl.py Normal file
View File

File diff suppressed because it is too large Load Diff

15
src/diffusers/utils/dummy_pt_objects.py
View File

@@ -362,6 +362,21 @@ class SD3Transformer2DModel(metaclass=DummyObject):
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"]

15
src/diffusers/utils/dummy_torch_and_transformers_objects.py
View File

@@ -107,6 +107,21 @@ class AnimateDiffSDXLPipeline(metaclass=DummyObject):
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"]

478
tests/pipelines/animatediff/test_animatediff_sparsectrl.py Normal file
View File

@@ -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)