LivePortrait
Open-source efficient portrait animation with stitching and retargeting control.
About
LivePortrait animates still portraits — humans, and even cats and dogs — by transferring expression and head pose from a driving video or motion template, using an implicit-keypoint framework instead of heavy diffusion sampling. Built by Kuaishou Technology with USTC and Fudan University researchers, it emphasizes controllability: explicit eye and lip retargeting plus a stitching module that blends the animated face back into the original image.
It runs fast enough for near-real-time use (ONNX-optimized), ships with a Gradio UI and ComfyUI support, and is MIT-licensed — a combination that made it one of the most-deployed open portrait animators (10k+ stars), including production use across Kuaishou's own apps. Limitations: quality depends heavily on the driving video, and extreme poses or occlusions can break the warp.
What it can do
- Single-image portrait animation from driving video
- Explicit eye and lip retargeting control
- Stitching module for seamless face blending
- Animal portrait mode and reusable motion templates
Setup
Clone https://github.com/KwaiVGI/LivePortrait, create a conda env, install PyTorch plus requirements, download pretrained weights, launch the Gradio app.
Prerequisites
Python 3.10, CUDA GPU, ~2GB disk for model weights.
Use cases
Works with
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 63; Trending score 75/100 from researcher assessment.
★ 10,000 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.