first-order-model

Why we included this project

This is the code behind the First Order Motion Model paper, and it remains a common way to get started with image animation. The core trick is that it transfers motion from a driving video onto a single source image, so expressions, head poses, and body movements carry over without any manual annotation or 3D reconstruction, which makes it a natural fit for talking-head demos and character animation experiments. Pretrained checkpoints cover faces, fashion, and other datasets, and a demo script plus Colab notebook get you from a driving clip and source image to an output video in a few commands. If you want to go further, the training and evaluation code is there too, so you can retrain on your own data. The codebase is compact enough to adapt for early avatar or motion-retargeting prototypes, even though newer methods have since moved past it.