VideoMaMa++: Temporally Consistent Video Matting via Preserve-and-Refine Token

Researchers introduced VideoMaMa++, a diffusion-based video matting method that uses a preserve-and-refine token to maintain temporal consistency across frames. The approach addresses the heavy compute of diffusion backbones, which the researchers say limits inference to short chunks, by leveraging video diffusion priors from internet-scale pretraining for zero-shot generalization from limited synthetic data.

Diffusion-based video matting has recently emerged as a leading approach, leveraging video diffusion priors from internet-scale pretraining to achieve strong zero-shot generalization from limited synthetic data. However, the heavy compute of these diffusion backbones limits inference to a short chun