MOFA-Video

Why we included this project

MOFA-Video is a research codebase from an ECCV 2024 paper that animates a single still image into a short clip. Rather than relying on a text prompt alone, it adapts a frozen image-to-video diffusion model with generative motion fields, so you can drive the animation with trajectory, keypoint, or hybrid controls. The repo ships Gradio demos and checkpoints for each control type, letting you test them side by side before committing. Training scripts for the trajectory variant are included, giving teams a concrete starting point for fine-tuning on their own data. It is research code, so expect to work with the provided scripts and checkpoints rather than a turnkey product, but the structure is clear enough to adapt for character animation, talking-head, or object-motion use cases.