update 🔥🔥🔥 We propose a face reenactment method, based on our AnimateAnyone pipeline: Using the facial landmark of driving video to control the pose of given source image, and keeping the identity of source image. Specially, we disentangle head attitude (including eyes blink) and mouth motion from the landmark of driving video, and it can control the expression and movements of source face precisely. We release our inference codes and pretrained models of face reenactment!!
update 🏋️🏋️🏋️ We release our training codes!! Now you can train your own AnimateAnyone models. See here for more details. Have fun!
update：🔥🔥🔥 We launch a HuggingFace Spaces demo of Moore-AnimateAnyone at here!!
This repository reproduces AnimateAnyone. To align the results demonstrated by the original paper, we adopt various approaches and tricks, which may differ somewhat from the paper and another implementation.
It's worth noting that this is a very preliminary version, aiming for approximating the performance (roughly 80% under our test) showed in AnimateAnyone.
We will continue to develop it, and also welcome feedbacks and ideas from the community. The enhanced version will also be launched on our MoBi MaLiang AIGC platform, running on our own full-featured GPU S4000 cloud computing platform.
- Inference codes and pretrained weights of AnimateAnyone
- Training scripts of AnimateAnyone
- Inference codes and pretrained weights of face reenactment
- Training scripts of face reenactment
- Inference scripts of audio driven portrait video generation
- Training scripts of audio driven portrait video generation
Here are some AnimateAnyone results we generated, with the resolution of 512x768.
compare-1-1.mp4
compare-2-2.mp4
demo3.mp4 |
demo4.mp4 |
demo5.mp4 |
demo6.mp4 |
Limitation: We observe following shortcomings in current version:
- The background may occur some artifacts, when the reference image has a clean background
- Suboptimal results may arise when there is a scale mismatch between the reference image and keypoints. We have yet to implement preprocessing techniques as mentioned in the paper.
- Some flickering and jittering may occur when the motion sequence is subtle or the scene is static.
These issues will be addressed and improved in the near future. We appreciate your anticipation!
Here are some results we generated, with the resolution of 512x512.
1.mp4 |
2.mp4 |
3.mp4 |
4.mp4 |
We Recommend a python version >=3.10 and cuda version =11.7. Then build environment as follows:
# [Optional] Create a virtual env
python -m venv .venv
source .venv/bin/activate
# Install with pip:
pip install -r requirements.txt
# For face landmark extraction
git clone https://github.com/emilianavt/OpenSeeFace.git Automatically downloading: You can run the following command to download weights automatically:
python tools/download_weights.pyWeights will be placed under the ./pretrained_weights direcotry. The whole downloading process may take a long time.
Manually downloading: You can also download weights manually, which has some steps:
-
Download our AnimateAnyone trained weights, which include four parts:
denoising_unet.pth,reference_unet.pth,pose_guider.pthandmotion_module.pth. -
Download our trained weights of face reenactment, and place these weights under
pretrained_weights. -
Download pretrained weight of based models and other components:
-
Download dwpose weights (
dw-ll_ucoco_384.onnx,yolox_l.onnx) following this.
Finally, these weights should be orgnized as follows:
./pretrained_weights/
|-- DWPose
| |-- dw-ll_ucoco_384.onnx
| `-- yolox_l.onnx
|-- image_encoder
| |-- config.json
| `-- pytorch_model.bin
|-- denoising_unet.pth
|-- motion_module.pth
|-- pose_guider.pth
|-- reference_unet.pth
|-- sd-vae-ft-mse
| |-- config.json
| |-- diffusion_pytorch_model.bin
| `-- diffusion_pytorch_model.safetensors
|-- reenact
| |-- denoising_unet.pth
| |-- reference_unet.pth
| |-- pose_guider1.pth
| |-- pose_guider2.pth
`-- stable-diffusion-v1-5
|-- feature_extractor
| `-- preprocessor_config.json
|-- model_index.json
|-- unet
| |-- config.json
| `-- diffusion_pytorch_model.bin
`-- v1-inference.yaml
Note: If you have installed some of the pretrained models, such as StableDiffusion V1.5, you can specify their paths in the config file (e.g. ./config/prompts/animation.yaml).
Here is the cli command for running inference scripts:
python -m scripts.pose2vid --config ./configs/prompts/animation.yaml -W 512 -H 784 -L 64You can refer the format of animation.yaml to add your own reference images or pose videos. To convert the raw video into a pose video (keypoint sequence), you can run with the following command:
python tools/vid2pose.py --video_path /path/to/your/video.mp4Here is the cli command for running inference scripts:
python -m scripts.lmks2vid --config ./configs/prompts/inference_reenact.yaml --driving_video_path YOUR_OWN_DRIVING_VIDEO_PATH --source_image_path YOUR_OWN_SOURCE_IMAGE_PATH We provide some face images in ./config/inference/talkinghead_images, and some face videos in ./config/inference/talkinghead_videos for inference.
Note: package dependencies have been updated, you may upgrade your environment via pip install -r requirements.txt before training.
Extract keypoints from raw videos:
python tools/extract_dwpose_from_vid.py --video_root /path/to/your/video_dirExtract the meta info of dataset:
python tools/extract_meta_info.py --root_path /path/to/your/video_dir --dataset_name anyone Update lines in the training config file:
data:
meta_paths:
- "./data/anyone_meta.json"Put openpose controlnet weights under ./pretrained_weights, which is used to initialize the pose_guider.
Put sd-image-variation under ./pretrained_weights, which is used to initialize unet weights.
Run command:
accelerate launch train_stage_1.py --config configs/train/stage1.yamlPut the pretrained motion module weights mm_sd_v15_v2.ckpt (download link) under ./pretrained_weights.
Specify the stage1 training weights in the config file stage2.yaml, for example:
stage1_ckpt_dir: './exp_output/stage1'
stage1_ckpt_step: 30000 Run command:
accelerate launch train_stage_2.py --config configs/train/stage2.yamlHuggingFace Demo: We launch a quick preview demo of Moore-AnimateAnyone at HuggingFace Spaces!! We appreciate the assistance provided by the HuggingFace team in setting up this demo.
To reduce waiting time, we limit the size (width, height, and length) and inference steps when generating videos.
If you have your own GPU resource (>= 16GB vram), you can run a local gradio app via following commands:
python app.py
- Installation for Windows users: Moore-AnimateAnyone-for-windows
We will launched this model on our MoBi MaLiang AIGC platform, running on our own full-featured GPU S4000 cloud computing platform. Mobi MaLiang has now integrated various AIGC applications and functionalities (e.g. text-to-image, controllable generation...). You can experience it by clicking this link or scanning the QR code bellow via WeChat!
This project is intended for academic research, and we explicitly disclaim any responsibility for user-generated content. Users are solely liable for their actions while using the generative model. The project contributors have no legal affiliation with, nor accountability for, users' behaviors. It is imperative to use the generative model responsibly, adhering to both ethical and legal standards.
We first thank the authors of AnimateAnyone. Additionally, we would like to thank the contributors to the majic-animate, animatediff and Open-AnimateAnyone repositories, for their open research and exploration. Furthermore, our repo incorporates some codes from dwpose and animatediff-cli-prompt-travel, and we extend our thanks to them as well.