Tencent-Hunyuan

HunyuanVideo-Avatar: Create Audio-Driven Character Videos

HunyuanVideo-Avatar generates dynamic, emotion-controllable videos of one or more characters from avatar images and audio. It is aimed at creators and researchers who need expressive talking-avatar or dialogue video generation and have access to compatible NVIDIA GPU hardware.

View on GitHub Compare
Stars
2.2k
Forks
349
Stars, last 30 days
+40

Our take

Health score

21/100

How it is scored
Activity
19
Community
22
Issues
21
Pull requests
20

+40 stars in the last 30 days

Project history

Daily GitHub snapshots since Oct 7, 2026. History covers stars and forks; health signals are not tracked historically.

Selected snapshot

Stars

2,221

Forks

351

Change since Nov 14, 2026: +2 stars·0 forks

Overview

HunyuanVideo-Avatar is a diffusion-transformer system for generating human and character animation from reference images and audio. It targets challenges in audio-driven video generation, including keeping a character visually consistent during motion, aligning expressions with emotional cues, and animating multiple characters in dialogue.

The project is most relevant to researchers and video creators exploring expressive avatar generation. Its inference code and model weights are available, with a project page, online playground, and Hugging Face model. The authors describe the method in their paper.

Key Features

  • Generates high-motion videos conditioned on avatar images and audio.
  • Supports photorealistic, cartoon, 3D-rendered, and anthropomorphic characters.
  • Handles portrait, upper-body, and full-body image scales.
  • Provides emotion control using cues from a reference image and target audio.
  • Supports multi-character dialogue, with audio directed to a selected character.
  • Includes single-GPU and multi-GPU inference paths, plus a Gradio server script.
  • Offers a low-memory inference route; the README also points to a TeaCache-based 10 GB VRAM option through Wan2GP.

Use Cases

  • Video creators can produce talking-avatar clips for social media or online content when they have a character image and audio track.
  • E-commerce and streaming teams can prototype expressive presenter videos without recording each animated performance.
  • Researchers can investigate audio-conditioned motion, emotion alignment, or multi-character animation using the released inference code and weights.
  • Editors and creators can generate dialogue scenes involving multiple animated characters from audio inputs.

Getting Started

See the README for model downloads, environment setup, and inference options. The documented setup begins by cloning the repository:

git clone https://github.com/Tencent-Hunyuan/HunyuanVideo-Avatar.git
cd HunyuanVideo-Avatar

Alternatives

  • InfiniteTalk: InfiniteTalk drives a person in an image or video with audio, while HunyuanVideo-Avatar focuses on expressive, emotion-controllable videos from avatar images and audio.
  • LivePortrait: LivePortrait animates portraits from a driving video or motion template, rather than generating audio-driven, emotion-controllable avatar videos.
ProjectLanguageLicenseStarsStatus
HunyuanVideo-AvatarPythonOther2.2kNot checked yet
InfiniteTalkPythonApache-2.08kQuiet
LivePortraitPythonOther19.2kQuiet

Considerations

  • Inference requires an NVIDIA GPU with CUDA support. The README gives 24 GB VRAM as a minimum for its stated configuration and recommends 96 GB for better generation quality; it also documents lower-memory alternatives.
  • Setup involves PyTorch, CUDA-compatible dependencies, pretrained model downloads, and FlashAttention for acceleration. Expect a substantially more involved setup than a typical Python utility.
  • The open-source plan lists inference and checkpoints as released, while ComfyUI support is marked as not yet released in the supplied README.
  • The repository metadata reports NOASSERTION for its license. Review the repository and model-weight terms before relying on them in a project.
  • The README includes a hosted playground and model links, but self-hosted inference still depends on suitable hardware and downloaded checkpoints.

Found this useful?

Share it with someone who would like HunyuanVideo-Avatar.

Comparisons