GaVS is an AI video stabilization tool that turns unstable footage into smoother, more cinematic-looking video. It uses 3D reconstruction to estimate camera motion and correct shake, making it suitable for clips with heavy vibration or complex scenes.
What GaVS does
GaVS stabilizes video with a 3D modeling and local reconstruction approach. It analyzes camera movement, compensates for jitter, and re-renders smoother frames while preserving the full frame (no edge cropping).
Key capabilities include:
- Stabilization for very shaky footage
- 3D scene reconstruction for more accurate motion estimation
- Rolling-shutter (“jello”) compensation
- Handling dynamic objects in the scene
- Full-frame stabilization without losing borders
- Adjustable stabilization strength via parameters
- Integration with monocular depth and optical flow
How to use
GaVS is available as a free desktop program via a GitHub repository. It runs locally and requires Python 3.10 and CUDA 12.6.
Typical setup steps:
- Install Python 3.10
- Download the GaVS repository from GitHub
- Create a conda virtual environment
- Install dependencies from requirements-torch.txt
- Download the dataset and pretrained models
- Run train.py to process video
- Tune stabilization settings in the config
- Run evaluate.py to check results
Documentation and the interface are in English.
