Tencent opened source code of HunyuanPortrait for photo animation

Tencent company published source code and weight coefficients of its HunyuanPortrait project – a system for creating realistic animated portraits. This solution, based on diffusion models, allows transferring movements from source video to static image, essentially “bringing to life” photographs with impressive degree of realism.

The technical architecture of HunyuanPortrait represents a comprehensive ensemble of specialized models. SVD for generating video sequences. DiNOv2 for extracting visual features. Arc2Face for working with facial expressions. And YoloFace for reliable face detection. Such multi-component structure provides more precise control over animation and smoothness of transitions between frames compared to existing analogues.

Current implementation requires 24 gigabytes of video memory to run the inference process, which makes the technology accessible to a relatively wide circle of enthusiasts and professionals having modern graphics accelerators at their disposal. It’s important to note that the system does not yet support “text-to-motion” mode, limiting itself to transferring movements from reference video.

Open access to code and model weights opens a new chapter in the development of digital animation technologies, potentially democratizing tools previously available only to large studios and technology giants.