AliaksandrSiarohin/first-order-model

Leverages keypoint-based motion transfer to animate still images using driving video sequences, with support for both absolute and relative coordinate modes to balance flexibility against pose consistency. The architecture uses a combination of keypoint detection, local affine transformations, and neural rendering, trained separately for domain-specific tasks (facial animation on VoxCeleb, fashion, animal motion). Includes pre-trained checkpoints, Docker/Colab deployment options, and integrations with face-alignment and motion co-segmentation libraries for extended capabilities like unsupervised face-swapping.

15,007 stars. No commits in the last 6 months.