StableAvatar
VerifiedOpen-source model that generates long talking avatar videos from a photo and audio
Product preview
StableAvatar
What it is
StableAvatar is an open-source research model for audio-driven avatar video generation, developed by researchers from Fudan University, Microsoft Research Asia, Xi'an Jiaotong University and Tencent. Given one reference image and an audio clip, it produces a talking or singing video with synchronized lips and natural motion, and it is designed to keep identity and quality stable over very long clips without face restoration tools. It builds on Wan2.1-1.3B with a time-step-aware audio adapter and a sliding-window strategy for smooth transitions. It supports 512x512, 480x832 and 832x480 output, runs through a Gradio app or ComfyUI, and includes LoRA fine-tuning code under the MIT license.
Key features
- Audio-driven avatar video from one image
- Long video generation without post-processing
- Gradio app and ComfyUI support
- LoRA and full fine-tuning code
What we liked
- + Free under MIT license
- + Low-memory offload mode available
- + Stable identity over long clips
Worth noting
- · Slow on consumer GPUs
- · Requires Python and model setup
Editor's notes
More context from AI BirdView
Many talking-head models look convincing for a few seconds and then drift. StableAvatar was built to tackle exactly that problem, and the authors explain that earlier approaches let audio errors accumulate across segments. Their audio adapter and sliding-window design aim to keep faces and lip sync stable across long videos, and the paper is listed as accepted at NeurIPS 2026.
- Generates video from a single reference image and an audio track
- Designed for very long clips without tools like FaceFusion or GFP-GAN
- Three supported resolutions in square, portrait and landscape layouts
- Gradio interface, ComfyUI support and a Hugging Face demo
- Full and LoRA fine-tuning scripts for customizing on your own data
StableAvatar is best for researchers, developers and technical creators who want a free, self-hosted way to make talking avatar videos. A practical tip: if your GPU has limited memory, use the sequential CPU offload mode, which the project says brings usage down to roughly 3 GB at the cost of speed, and test with short clips before rendering longer ones.
Reader reviews
From people who've used StableAvatar. Emails are never shown publicly.
No reviews yet. Be the first to share your experience.
Alternatives to StableAvatar
Other tools we like in this space.
Hedra
Visual AI studio and infrastructure for character-driven video
Sync
State-of-the-art lip sync for any video, any language
OmniHuman 2.0
Turn Images and Audio into Lifelike AI Videos