- Home
- AI Video Lip Sync Tools
- KeySync
On the recordRead 9 Oct 202621 facts
KeySync
On the record
- Architecture
- KeySync is a two-stage framework designed to improve temporal consistency in lip synchronization.antonibigata.github.io · read 9 Oct 2026
- Demo limit
- The Hugging Face demo is limited to videos of up to 6 seconds because of GPU restrictions.github.com · read 9 Oct 2026
- Evaluation
- The repository includes a LipScore metric for measuring synchronization quality between generated and ground-truth videos.github.com · read 9 Oct 2026
- Expression leakage
- Its masking strategy is designed to reduce expression leakage from the input video.antonibigata.github.io · read 4 Oct 2026
- Framework
- KeySync uses a two-stage framework to improve temporal consistency in lip synchronization.antonibigata.github.io · read 4 Oct 2026
- Input formats
- The quick start guide uses MP4 video files and WAV audio files as inputs.github.com · read 4 Oct 2026
- Input preparation
- The repository recommends converting video to 25 fps and audio to 16000 Hz for compatibility with the models.github.com · read 9 Oct 2026
- Interactive demo
- The demo accepts video and audio uploads to create synchronized videos and is limited to videos up to 6 seconds by Hugging Face Spaces GPU restrictions.github.com · read 4 Oct 2026
- Leakage and occlusion handling
- The framework uses a masking strategy to address expression leakage and facial occlusions.antonibigata.github.io · read 9 Oct 2026
- License
- The GitHub repository identifies its license as Apache-2.0.github.com · read 9 Oct 2026
- Local requirements
- The repository lists a CUDA-compatible GPU, Python 3.11, and Conda as prerequisites for local use.github.com · read 9 Oct 2026
- Model weights
- The repository instructs users to download pretrained models from the toninio19/keysync Hugging Face repository.github.com · read 4 Oct 2026
- Occlusion controls
- The command-line options include a setting to enable occlusion handling and coordinates for the object to mask.github.com · read 4 Oct 2026
- Occlusion option
- The inference parameters include an option to mask objects that block the face.github.com · read 9 Oct 2026
- Occlusions
- The framework includes a masking approach for handling facial occlusions.antonibigata.github.io · read 4 Oct 2026
- Online demo
- The interactive demo accepts video and audio uploads to create synchronized videos.github.com · read 9 Oct 2026
- Preprocessing
- Provided scripts can convert videos to 25 fps and audio to 16000 Hz, and crop videos around the face.github.com · read 4 Oct 2026
- Pretrained models
- The repository provides instructions to download pretrained models from Hugging Face.github.com · read 9 Oct 2026
