MuseV English 中文

MuseV: Infinite-length and High Fidelity Virtual Human Video Generation with Visual Conditioned Parallel Denoising
Zhiqiang Xia ^*, Zhaokang Chen^*, Bin Wu^†, Chao Li, Kwok-Wai Hung, Chao Zhan, Yingjie He, Wenjiang Zhou (^*co-first author, ^†Corresponding Author, [email protected])

github huggingface HuggingfaceSpace project Technical report (comming soon)

We have setup the world simulator vision since March 2023, believing diffusion models can simulate the world. MuseV was a milestone achieved around July 2023. Amazed by the progress of Sora, we decided to opensource MuseV, hopefully it will benefit the community. Next we will move on to the promising diffusion+transformer scheme.

Update: We have released MuseTalk, a real-time high quality lip sync model, which can be applied with MuseV as a complete virtual human generation solution.

Overview

MuseV is a diffusion-based virtual human video generation framework, which

supports infinite length generation using a novel Visual Conditioned Parallel Denoising scheme.
checkpoint available for virtual human video generation trained on human dataset.
supports Image2Video, Text2Image2Video, Video2Video.
compatible with the Stable Diffusion ecosystem, including base_model, lora, controlnet, etc.
supports multi reference image technology, including IPAdapter, ReferenceOnly, ReferenceNet, IPAdapterFaceID.
training codes (comming very soon).

Important bug fixes

musev_referencenet_pose: model_name of unet, ip_adapter of Command is not correct, please use musev_referencenet_pose instead of musev_referencenet.

News

[03/27/2024] release MuseV project and trained model musev, muse_referencenet.
[03/30/2024] add huggingface space gradio to generate video in gui

Model

Overview of model structure

Parallel denoising

Cases

All frames were generated directly from text2video model, without any post process.

Examples bellow can be accessed at configs/tasks/example.yaml MoreCase is in project

Text/Image2Video

Human

image	video	prompt
		(masterpiece, best quality, highres:1),(1boy, solo:1),(eye blinks:1.8),(head wave:1.3)
		(masterpiece, best quality, highres:1),(1girl, solo:1),(beautiful face, soft skin, costume:1),(eye blinks:{eye_blinks_factor}),(head wave:1.3)
		(masterpiece, best quality, highres:1), peaceful beautiful sea scene
		(masterpiece, best quality, highres:1), peaceful beautiful sea scene
		(masterpiece, best quality, highres:1), playing guitar
		(masterpiece, best quality, highres:1), playing guitar
		(masterpiece, best quality, highres:1), playing guitar
		(masterpiece, best quality, highres:1), playing guitar
		(masterpiece, best quality, highres:1),(1man, solo:1),(eye blinks:1.8),(head wave:1.3),Chinese ink painting style
		(masterpiece, best quality, highres:1),(1girl, solo:1),(beautiful face, soft skin, costume:1),(eye blinks:{eye_blinks_factor}),(head wave:1.3)
		(masterpiece, best quality, highres:1),(1man, solo:1),(eye blinks:1.8),(head wave:1.3)
		(masterpiece, best quality, highres:1),(1man, solo:1),(eye blinks:1.8),(head wave:1.3), animate
		(masterpiece, best quality, highres:1),(1girl, solo:1),(beautiful face, soft skin, costume:1),(eye blinks:{eye_blinks_factor}),(head wave:1.3)

Scene

image	video	prompt
		(masterpiece, best quality, highres:1), peaceful beautiful waterfall, an endless waterfall
		(masterpiece, best quality, highres:1), peaceful beautiful river
		(masterpiece, best quality, highres:1), peaceful beautiful sea scene

VideoMiddle2Video

pose2video In duffy mode, pose of the vision condition frame is not aligned with the first frame of control video. posealign will solve the problem.

image	video	prompt
		(masterpiece, best quality, highres:1) , a girl is dancing, animation
		(masterpiece, best quality, highres:1), is dancing, animation

MuseTalk

The character of talk, Sun Xinying is a supermodel KOL. You can follow her on douyin.

name	video
talk
talk
sing

TODO:

technical report (comming soon).
training codes.
release pretrained unet model, which is trained with controlnet、referencenet、IPAdapter, which is better on pose2video.
support diffusion transformer generation framework.
release posealign module

Quickstart

Prepare python environment and install extra package like diffusers, controlnet_aux, mmcm.

Prepare environment

You are recommended to use docker primarily to prepare python environment.

prepare python env

Attention: we only test with docker, there are maybe trouble with conda, or requirement. We will try to fix it. Use docker Please.

Method 1: docker

pull docker image

docker pull anchorxia/musev:latest

run docker

docker run --gpus all -it --entrypoint /bin/bash anchorxia/musev:latest

The default conda env is musev.

Method 2: conda

create conda environment from environment.yaml

conda env create --name musev --file ./environment.yml

Method 3: pip requirements

pip install -r requirements.txt

Prepare mmlab package

if not use docker, should install mmlab package additionally.

pip install --no-cache-dir -U openmim 
mim install mmengine 
mim install "mmcv>=2.0.1" 
mim install "mmdet>=3.1.0" 
mim install "mmpose>=1.1.0"

Prepare custom package / modified package

clone

git clone --recursive https://github.com/TMElyralab/MuseV.git

prepare PYTHONPATH

current_dir=$(pwd)
export PYTHONPATH=${PYTHONPATH}:${current_dir}/MuseV
export PYTHONPATH=${PYTHONPATH}:${current_dir}/MuseV/MMCM
export PYTHONPATH=${PYTHONPATH}:${current_dir}/MuseV/diffusers/src
export PYTHONPATH=${PYTHONPATH}:${current_dir}/MuseV/controlnet_aux/src
cd MuseV

MMCM: multi media, cross modal process package。
diffusers: modified diffusers package based on diffusers
controlnet_aux: modified based on controlnet_aux

Download models

git clone https://huggingface.co./TMElyralab/MuseV ./checkpoints

motion: text2video model, trained on tiny ucf101 and tiny webvid dataset, approximately 60K videos text pairs. GPU memory consumption testing on resolution$=512*512$, time_size=12.
- musev/unet: only has and train unet motion module. GPU memory consumption $\approx 8G$.
- musev_referencenet: train unet module, referencenet, IPAdapter. GPU memory consumption $\approx 12G$.
  - unet: motion module, which has to_k, to_v in Attention layer refer to IPAdapter
  - referencenet: similar to AnimateAnyone
  - ip_adapter_image_proj.bin: images clip emb project layer, refer to IPAdapter
- musev_referencenet_pose: based on musev_referencenet, fix referencenetand controlnet_pose, train unet motion and IPAdapter. GPU memory consumption $\approx 12G$
t2i/sd1.5: text2image model, parameter are frozen when training motion module. Different t2i base_model has a significant impact.
- majicmixRealv6Fp16: example, could be replaced with other t2i base. download from majicmixRealv6Fp16
- fantasticmix_v10: example, could be replaced with other t2i base. download from fantasticmix_v10
IP-Adapter/models: download from IPAdapter
- image_encoder: vision clip model.
- ip-adapter_sd15.bin: original IPAdapter model checkpoint.
- ip-adapter-faceid_sd15.bin: original IPAdapter model checkpoint.

Inference

Prepare model_path

Skip this step when run example task with example inference command. Set model path and abbreviation in config, to use abbreviation in inference script.

T2I SD：ref to musev/configs/model/T2I_all_model.py
Motion Unet: refer to musev/configs/model/motion_model.py
Task: refer to musev/configs/tasks/example.yaml

musev_referencenet

text2video

python scripts/inference/text2video.py   --sd_model_name majicmixRealv6Fp16   --unet_model_name musev_referencenet --referencenet_model_name musev_referencenet --ip_adapter_model_name musev_referencenet   -test_data_path ./configs/tasks/example.yaml  --output_dir ./output  --n_batch 1  --target_datas yongen  --vision_clip_extractor_class_name ImageClipVisionFeatureExtractor --vision_clip_model_path ./checkpoints/IP-Adapter/models/image_encoder  --time_size 12 --fps 12

common parameters:

test_data_path: task_path in yaml extention
target_datas: sep is ,, sample subtasks if name in test_data_path is in target_datas.
sd_model_cfg_path: T2I sd models path, model config path or model path.
sd_model_name: sd model name, which use to choose full model path in sd_model_cfg_path. multi model names with sep =,, or all
unet_model_cfg_path: motion unet model config path or model path。
unet_model_name: unet model name, use to get model path in unet_model_cfg_path, and init unet class instance in musev/models/unet_loader.py. multi model names with sep=,, or all. If unet_model_cfg_path is model path, unet_name must be supported in musev/models/unet_loader.py
time_size: num_frames per diffusion denoise generation。default=12.
n_batch: generation numbers of shot, $total_frames=n_batch * time_size + n_viscond$, default=1。
context_frames: context_frames num. If time_size > context_frame，time_size window is split into many sub-windows for parallel denoising"。 default=12。

To generate long videos, there two ways:

visual conditioned parallel denoise: set n_batch=1, time_size = all frames you want.
traditional end-to-end: set time_size = context_frames = frames of a shot (12), context_overlap = 0；

model parameters： supports referencenet, IPAdapter, IPAdapterFaceID, Facein.

referencenet_model_name: referencenet model name.
ImageClipVisionFeatureExtractor: ImageEmbExtractor name, extractor vision clip emb used in IPAdapter.
vision_clip_model_path: ImageClipVisionFeatureExtractor model path.
ip_adapter_model_name: from IPAdapter, it's ImagePromptEmbProj, used with ImageEmbExtractor。
ip_adapter_face_model_name: IPAdapterFaceID, from IPAdapter to keep faceid，should set face_image_path。

Some parameters that affect the motion range and generation results：

video_guidance_scale: Similar to text2image, control influence between cond and uncond，default=3.5
guidance_scale: The parameter ratio in the first frame image between cond and uncond, default=3.5
use_condition_image: Whether to use the given first frame for video generation.
redraw_condition_image: Whether to redraw the given first frame image.
video_negative_prompt: Abbreviation of full negative_prompt in config path. default=V2.

video2video

t2i base_model has a significant impact. In this case, fantasticmix_v10 performs better than majicmixRealv6Fp16.

python scripts/inference/video2video.py --sd_model_name fantasticmix_v10  --unet_model_name musev_referencenet --referencenet_model_name   musev_referencenet --ip_adapter_model_name musev_referencenet    -test_data_path ./configs/tasks/example.yaml    --vision_clip_extractor_class_name ImageClipVisionFeatureExtractor --vision_clip_model_path ./checkpoints/IP-Adapter/models/image_encoder      --output_dir ./output  --n_batch 1 --controlnet_name dwpose_body_hand  --which2video "video_middle"  --target_datas dance1 --fps 12 --time_size 12

import parameters

Most of the parameters are same as musev_text2video. Special parameters of video2video are:

need to set video_path in test_data. Now supports rgb video and controlnet_middle_video。

which2video: whether rgb video influences initial noise, more strongly than controlnet condition. If True, then redraw video.
controlnet_name：whether to use controlnet condition, such as dwpose,depth.
video_is_middle: video_path is rgb video or controlnet_middle_video. Can be set for every test_data in test_data_path.
video_has_condition: whether condtion_images is aligned with the first frame of video_path. If Not, firstly generate condition_images and then align with concatation. set in test_data。

all controlnet_names refer to mmcm

['pose', 'pose_body', 'pose_hand', 'pose_face', 'pose_hand_body', 'pose_hand_face', 'dwpose', 'dwpose_face', 'dwpose_hand', 'dwpose_body', 'dwpose_body_hand', 'canny', 'tile', 'hed', 'hed_scribble', 'depth', 'pidi', 'normal_bae', 'lineart', 'lineart_anime', 'zoe', 'sam', 'mobile_sam', 'leres', 'content', 'face_detector']

musev_referencenet_pose

Only used for pose2video train based on musev_referencenet, fix referencenet, pose-controlnet, and T2I, train motion module and IPAdapter.

t2i base_model has a significant impact. In this case, fantasticmix_v10 performs better than majicmixRealv6Fp16.

python scripts/inference/video2video.py --sd_model_name fantasticmix_v10  --unet_model_name musev_referencenet_pose --referencenet_model_name   musev_referencenet --ip_adapter_model_name musev_referencenet_pose    -test_data_path ./configs/tasks/example.yaml    --vision_clip_extractor_class_name ImageClipVisionFeatureExtractor --vision_clip_model_path ./checkpoints/IP-Adapter/models/image_encoder      --output_dir ./output  --n_batch 1 --controlnet_name dwpose_body_hand  --which2video "video_middle"  --target_datas  dance1   --fps 12 --time_size 12

musev

Only has motion module, no referencenet, requiring less gpu memory.

text2video

python scripts/inference/text2video.py   --sd_model_name majicmixRealv6Fp16   --unet_model_name musev   -test_data_path ./configs/tasks/example.yaml  --output_dir ./output  --n_batch 1  --target_datas yongen  --time_size 12 --fps 12

video2video

python scripts/inference/video2video.py --sd_model_name fantasticmix_v10  --unet_model_name musev    -test_data_path ./configs/tasks/example.yaml --output_dir ./output  --n_batch 1 --controlnet_name dwpose_body_hand  --which2video "video_middle"  --target_datas  dance1   --fps 12 --time_size 12

Gradio demo

MuseV provides gradio script to generate a GUI in a local machine to generate video conveniently.

cd scripts/gradio
python app.py

Acknowledgements

MuseV has referred much to TuneAVideo, diffusers, Moore-AnimateAnyone, animatediff, IP-Adapter, AnimateAnyone, VideoFusion, insightface.
MuseV has been built on ucf101 and webvid datasets.

Thanks for open-sourcing!

Limitation

There are still many limitations, including

Lack of generalization ability. Some visual condition image perform well, some perform bad. Some t2i pretraied model perform well, some perform bad.
Limited types of video generation and limited motion range, partly because of limited types of training data. The released MuseV has been trained on approximately 60K human text-video pairs with resolution 512*320. MuseV has greater motion range while lower video quality at lower resolution. MuseV tends to generate less motion range with high video quality. Trained on larger, higher resolution, higher quality text-video dataset may make MuseV better.
Watermarks may appear because of webvid. A cleaner dataset withour watermarks may solve this issue.
Limited types of long video generation. Visual Conditioned Parallel Denoise can solve accumulated error of video generation, but the current method is only suitable for relatively fixed camera scenes.
Undertrained referencenet and IP-Adapter, beacause of limited time and limited resources.
Understructured code. MuseV supports rich and dynamic features, but with complex and unrefacted codes. It takes time to familiarize.

Citation

@article{musev,
  title={MuseV: Infinite-length and High Fidelity Virtual Human Video Generation with Visual Conditioned Parallel Denoising},
  author={Xia, Zhiqiang and Chen, Zhaokang and Wu, Bin and Li, Chao and Hung, Kwok-Wai and Zhan, Chao and He, Yingjie and Zhou, Wenjiang},
  journal={arxiv},
  year={2024}
}

Disclaimer/License

code: The code of MuseV is released under the MIT License. There is no limitation for both academic and commercial usage.
model: The trained model are available for non-commercial research purposes only.
other opensource model: Other open-source models used must comply with their license, such as insightface, IP-Adapter, ft-mse-vae, etc.
The testdata are collected from internet, which are available for non-commercial research purposes only.
AIGC: This project strives to impact the domain of AI-driven video generation positively. Users are granted the freedom to create videos using this tool, but they are expected to comply with local laws and utilize it responsibly. The developers do not assume any responsibility for potential misuse by users.