OPEN LINK ↗
// 1h agoRESEARCH PAPER
VideoMaMa++ fixes flickering video matte seams
VideoMaMa++ is a mask-guided diffusion model from KAIST, Korea University, and Adobe Research that preserves fine alpha details across long videos. Its preserve/refine tokens, sliding-window inference, and reference anchors reduce edge flicker and transparency jumps between processing chunks.
// ANALYSIS
VideoMaMa++ targets the practical failure mode that makes diffusion-based matting difficult to deploy: temporal instability across chunk boundaries.
- –Separate RGB and mask token streams improve fine-detail reconstruction.
- –Preserve tokens carry prior mattes forward, while refine tokens process new frames.
- –The model reportedly leads VideoMatte, YouTubeMatte, and V-HIM60 benchmarks.
- –It still depends on upstream mask tracking and substantial diffusion inference compute.
- –Code and weights remain under review, limiting immediate self-hosted experimentation.
// TAGS
videomama++visionmultimodalinferenceresearch
DISCOVERED
1h ago
2026-10-22
PUBLISHED
2h ago
2026-10-22
RELEVANCE
8/ 10
AUTHOR
cutetoxicguy