few-shot-vid2vid

Why we included this project

This repo is the NeurIPS 2019 release of NVIDIA's few-shot video-to-video translation method. Trained on a handful of example images of a target subject, it learns to render photorealistic video from guidance like human poses, face edge maps, or semantic label maps. The included scenarios cover the main use cases: dance transfer from pose keypoints, talking heads from facial landmarks, and converting street-view segmentation into realistic footage. It's a solid reference for how adaptive SPADE normalization, feature warping, and temporal smoothing work together in one pipeline, though NVIDIA has since deprecated the code in favor of Imaginaire, so it's best seen as a stable way to learn the method and reproduce the published results.