HunyuanVideo-Avatar

Part of Tencent's HunyuanVideo model family, HunyuanVideo-Avatar turns a character image and an audio track into a talking, emoting video using a multimodal diffusion transformer (MM-DiT). Three modules define the approach: a character image injection mechanism that keeps the avatar consistent across frames, an Audio Emotion Module that transfers emotional cues from a reference image, and a Face-Aware Audio Adapter that isolates audio per face so multiple characters in one scene can speak independently. It handles photorealistic, cartoon, 3D-rendered, and anthropomorphic styles at portrait through full-body scales, targeting e-commerce, streaming, and video editing. Inference officially calls for an NVIDIA GPU with 24 GB of VRAM and 96 GB for best quality, though a June 2025 TeaCache update runs on a single 10 GB card; setup is a conda environment with Python 3.10 and PyTorch 2.4, with Docker images available. Weights ship under the Tencent Hunyuan Community License, which excludes the EU, UK, and South Korea and requires a separate agreement above 100 million monthly users. The repository has about 2,100 stars.