KlingAIResearch /LivePortrait

About LivePortrait

LivePortrait animates a still face photo or video with motion from another clip. It's the official PyTorch code from Kuaishou, and it handles humans, cats and dogs. You get command-line scripts and a Gradio interface (a browser UI for the model). Animal mode needs Linux or Windows with an NVIDIA GPU.

Best for: Talking-head animators who need controllable face animation or expression transfer.

• Drives a source photo or video with motion from a separate driving video
• Animates cat and dog faces in Animals mode, built on X-Pose keypoints
• Edits portrait videos (v2v) and supports image-driven mode with regional control
• Lets you adjust pose and expression in the Gradio interface
• Saves motion as .pkl templates for faster, more private reuse
• Auto-crops driving videos and benchmarks speed for each module

1. 1
   Clone the repo and make a conda env: the command below, the next command, conda activate LivePortrait

   git clone https://github.com/KlingTeam/LivePortrait

   Copy

   conda create -n LivePortrait python=3.10

   Copy
2. 2
   Install PyTorch for your CUDA version, then the command below (macOS Apple Silicon: the next command)

   pip install -r requirements.txt

   Copy

   pip install -r requirements_macOS.txt

   Copy
3. 3
   Download the weights: the command below

   huggingface-cli download KlingTeam/LivePortrait --local-dir pretrained_weights --exclude "*.git*" "README.md" "docs"

   Copy
4. 4
   Run the human demo: python inference.py, or set your own inputs with the command below

   python inference.py -s assets/examples/source/s9.jpg -d assets/examples/driving/d0.mp4

   Copy
5. 5
   Open the Gradio UI with python app.py; for animals, build the MultiScaleDeformableAttention op first, then run python app_animals.py