This repository contains source code for RAIN: Real-time Animation Of Infinite Video Stream. A real-time implementation for video generataion on customer-level devices.
Project Page is here.
- Release Demo code
- Release inference pipeline code
- Release source code for training
- Update for a more interactive implementation
We recommend python >=3.10. Install pytorch on the official website firstly (We recommend torch >=2.3.0).
git clone https://github.com/Pscgylotti/RAIN.git
# clone repository from github
cd RAIN
pip install -r requirements_inference.txt
# install requirements for inferencing
You can download original RAIN weights from Google Drive, Huggingeface Hub, and then put them into weights/torch/.
You can get 'taesdv.pth' from https://github.com/madebyollin/taesdv, and put it into weights/torch/.
Clone https://huggingface.co/stabilityai/sd-vae-ft-mse into weights/.
Download 'image_encoder' folder and its contents from https://huggingface.co/lambdalabs/sd-image-variations-diffusers and put it into weights/.
Download 'dw-mm_ucoco.onnx' from https://huggingface.co/hr16/UnJIT-DWPose/tree/main and 'yolox_s.onnx' from https://github.com/Megvii-BaseDetection/YOLOX/releases/tag/0.1.1rc0, and put them into weights/onnx. (You can choose to use 'dw-ll_ucoco-384.onnx' and 'yolox_l.onnx' from https://huggingface.co/yzd-v/DWPose/tree/main for higher accuracy).
(You can always redirect the weights directory in configs/rain_morpher.yaml)
In configs/rain_morpher.yaml, modify tensorrt: False into tensorrt: True to enable TensorRT acceleration. In the first launch it will take about ten minutes to compile the model.
Simply execute python gradio_app.py and open http://localhost:7860/ in browser (Usually the port is 7860).
Upload an upper-half-body potrait of any anime characters, fuse reference and turn on the web camera. Click on start to launch face morphing. You may need to adjust some morphing parameters to fit your face with the character face (Especially for the eye-related parameters. The eyes generally fail to synthesize with incompatible parameters).
It generally takes about 12 GiB of Device RAM to run the whole inference demo. However, you can unload the reference part after fusing the reference image. Then the synthesis-only model requires about 8 GiB of Device RAM to run.
Special thanks to CivitAI Community and YODOYA for example images. Thanks to Jianwen Meng for pipeline design.