Our take
- Activity
- 19
- Community
- 43
- Issues
- 46
- Pull requests
- 20
+318 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.
Stars
8,391Forks
1,443Change since Nov 14, 2026: +3 stars·+1 forks
Overview
InfiniteTalk is a Python project for generating audio-driven talking videos from either a still image or an existing video. For video dubbing, it uses the audio to animate lip movement as well as head motion, posture, and facial expression, while seeking to preserve the person’s identity.
Its streaming mode is designed for long video generation, making it relevant when short-clip dubbing tools are not sufficient. The project includes inference code, model weights, a Gradio interface, and integration paths for ComfyUI and Wan2GP.
Key Features
- Supports both image-to-video generation and video-to-video dubbing.
- Synchronizes speech with lips, head movement, body posture, and facial expressions.
- Offers streaming mode for long video generation and clip mode for shorter output.
- Provides single-GPU and multi-GPU inference paths.
- Includes options for reduced memory use, including low-VRAM settings and an FP8 quantized model.
- Supports 480p and 720p output configurations.
- Offers a Gradio interface and community integrations for ComfyUI and Wan2GP.
Use Cases
- Video creators can dub an existing talking-head video to new audio while retaining its original camera movement.
- Teams producing narrated content can animate a still portrait from an audio track when they do not have source video.
- Developers can prototype audio-driven character animation or integrate generation into ComfyUI-based workflows.
- Researchers can explore sparse-frame video dubbing and audio-conditioned human video generation using the released code and model weights.
Getting Started
The README documents a CUDA-oriented setup using Python 3.10, PyTorch, xformers, FlashAttention, and FFmpeg. After installing dependencies and downloading the required base and project weights, a single-GPU inference run can be started with:
python generate_infinitetalk.py --ckpt_dir weights/Wan2.1-I2V-14B-480P --wav2vec_dir weights/chinese-wav2vec2-base --infinitetalk_dir weights/InfiniteTalk/single/infinitetalk.safetensors --input_json examples/single_example_image.json --size infinitetalk-480 --sample_steps 40 --mode streaming --motion_frame 9 --save_file infinitetalk_res
See the repository README for full installation, model download, multi-GPU, and interface instructions.
Alternatives
- audio2photoreal: audio2photoreal generates photorealistic conversational avatars from audio, while InfiniteTalk animates an image or video for audio-driven speech.
- MuseTalk: MuseTalk focuses on lip-synced video from a source image or video and audio, while InfiniteTalk also targets facial expressions and body movement in long-form videos.
| Project | Language | License | Stars | Status |
|---|---|---|---|---|
| InfiniteTalk | Python | Apache-2.0 | 8k | Not checked yet |
| audio2photoreal | Python | Other | 2.8k | Archived |
| MuseTalk | Python | Other | 6.7k | Inactive |
Considerations
- Inference depends on large pretrained model files, including the Wan2.1-I2V-14B base model, and the documented setup uses CUDA-specific packages. Plan for a capable GPU and substantial storage and memory; low-VRAM and quantized options are available but have their own constraints.
- The README reports that image-to-video color shifts become more pronounced beyond one minute. FusionX LoRA can also increase color shift over one minute and reduce identity preservation.
- Video-to-video generation can imitate, but not exactly reproduce, the source camera movement. SDEdit may improve camera matching for short clips, with a color-shift tradeoff.
- The repository offers implementation and model weights under Apache-2.0, but generated content remains subject to applicable laws and the usage responsibilities described by the project.
- Some planned acceleration work remains incomplete, including LCM distillation and sparse attention.
Project links: GitHub repository · Project page · Technical report · Model weights
Found this useful?
Share it with someone who would like InfiniteTalk.