VideoMaMa Mask-Guided Video Matting via Generative Prior

paper:

The edge consistency on the hair is honestly the most impressive part here—that’s usually where these models fall apart. The smart pivot with VideoMaMa is using a generative prior (based on Stable Video Diffusion) rather than just pixel classification. It basically takes a rough binary mask and "hallucinates" the fine structural details to create a perfect alpha matte. It bridges the gap between simple segmentation and actual professional matting. Also, the fact that they built the MA-V dataset (50k+ real-world videos) to train this might be even more valuable than the model itself. They are finally fixing the data scarcity bottleneck for video tasks. Curious to see how heavy this is to run locally compared to lighter tools like RVM.

When Comfy is all I care about

GM