mingkin / FasterLivePortrait

Bring portraits to life in Real Time！onnx/tensorrt support！

FasterLivePortrait: Bring portraits to life in Real Time!

English | 中文

Original repository: LivePortrait, thanks to the authors for sharing

New features:

• Achieved real-time running of LivePortrait on RTX 3090 GPU using TensorRT, reaching speeds of 30+ FPS. This is the speed for rendering a single frame, including pre- and post-processing, not just the model inference speed.
• Implemented conversion of LivePortrait model to Onnx model, achieving inference speed of about 70ms/frame (~12 FPS) using onnxruntime-gpu on RTX 3090, facilitating cross-platform deployment.
• Seamless support for native gradio app, with several times faster speed and support for simultaneous inference on multiple faces. Some results can be seen here: pr105
• Refactored code structure, no longer dependent on pytorch, all models use onnx or tensorrt for inference.

If you find this project useful, please give it a star ❤️❤️

result2.mp4

Changelog

• 2024/07/17: Added support for Docker environment, providing a runnable image.
• 2024/07/18: macOS support added(No need for Docker, Python is enough). M1/M2 chips are faster, but it's still quite slow 😟
  • Install ffmpeg: brew install ffmpeg
  • Set up a Python 3.10 virtual environment. Recommend using miniforge: conda create -n flip python=3.10 && conda activate flip
  • Install requirements: pip install -r requirements_macos.txt
  • Download ONNX files: huggingface-cli download warmshao/FasterLivePortrait --local-dir ./checkpoints
  • Test: python app.py --mode onnx
• Windows integration package, supports one-click run

Environment Setup

• Option 1: Docker (recommended).A docker image is provided for eliminating the need to install onnxruntime-gpu and TensorRT manually.
  • Install Docker according to your system
  • Download the image: docker pull shaoguo/faster_liveportrait:v1
  • Execute the command, replace $FasterLivePortrait_ROOT with the local directory where you downloaded FasterLivePortrait:

  docker run -it --gpus=all \
  --name faster_liveportrait \
  -v $FasterLivePortrait_ROOT:/root/FasterLivePortrait \
  --restart=always \
  -p 9870:9870 \
  shaoguo/faster_liveportrait:v1 \
  /bin/bash

• Option 2: Create a new Python virtual environment and install the necessary Python packages manually.
  • First, install ffmpeg
  • Run pip install -r requirements.txt
  • Then follow the tutorials below to install onnxruntime-gpu or TensorRT. Note that this has only been tested on Linux systems.

Onnxruntime Inference

• First, download the converted onnx model files:huggingface-cli download warmshao/FasterLivePortrait --local-dir ./checkpoints.
• (Ignored in Docker)If you want to use onnxruntime cpu inference, simply pip install onnxruntime. However, cpu inference is extremely slow and not recommended. The latest onnxruntime-gpu still doesn't support grid_sample cuda, but I found a branch that supports it. Follow these steps to install onnxruntime-gpu from source:
  • git clone https://github.com/microsoft/onnxruntime
  • git checkout liqun/ImageDecoder-cuda. Thanks to liqun for the grid_sample with cuda implementation!
  • Run the following commands to compile, changing cuda_version and CMAKE_CUDA_ARCHITECTURES according to your machine:

  ./build.sh --parallel \
  --build_shared_lib --use_cuda \
  --cuda_version 11.8 \
  --cuda_home /usr/local/cuda --cudnn_home /usr/local/cuda/ \
  --config Release --build_wheel --skip_tests \
  --cmake_extra_defines CMAKE_CUDA_ARCHITECTURES="60;70;75;80;86" \
  --cmake_extra_defines CMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
  --disable_contrib_ops \
  --allow_running_as_root

  • pip install build/Linux/Release/dist/onnxruntime_gpu-1.17.0-cp310-cp310-linux_x86_64.whl
• Test the pipeline using onnxruntime:

    python run.py \
   --src_image assets/examples/source/s10.jpg \
   --dri_video assets/examples/driving/d14.mp4 \
   --cfg configs/onnx_infer.yaml
  

TensorRT Inference

• (Ignored in Docker) Install TensorRT. Remember the installation path of TensorRT.
• (Ignored in Docker) Install the grid_sample TensorRT plugin, as the model uses grid sample that requires 5D input, which is not supported by the native grid_sample operator.
  • git clone https://github.com/SeanWangJS/grid-sample3d-trt-plugin
  • Modify line 30 in CMakeLists.txt to: set_target_properties(${PROJECT_NAME} PROPERTIES CUDA_ARCHITECTURES "60;70;75;80;86")
  • export PATH=/usr/local/cuda/bin:$PATH
  • mkdir build && cd build
  • cmake .. -DTensorRT_ROOT=$TENSORRT_HOME, replace $TENSORRT_HOME with your own TensorRT root directory.
  • make, remember the address of the .so file, replace /opt/grid-sample3d-trt-plugin/build/libgrid_sample_3d_plugin.so in scripts/onnx2trt.py and src/models/predictor.py with your own .so file path
• Download ONNX model files:huggingface-cli download warmshao/FasterLivePortrait --local-dir ./checkpoints. Convert all ONNX models to TensorRT, run sh scripts/all_onnx2trt.sh
• Test the pipeline using tensorrt:

   python run.py \
   --src_image assets/examples/source/s10.jpg \
   --dri_video assets/examples/driving/d14.mp4 \
   --cfg configs/trt_infer.yaml

• To run in real-time using a camera:

   python run.py \
   --src_image assets/examples/source/s10.jpg \
   --dri_video 0 \
   --cfg configs/trt_infer.yaml \
   --realtime

Gradio App

• onnxruntime: python app.py --mode onnx
• tensorrt: python app.py --mode trt
• The default port is 9870. Open the webpage: http://localhost:9870/

About Me

Follow my shipinhao channel for continuous updates on my AIGC content. Feel free to message me for collaboration opportunities.