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

Forum topic · 小凯 · 2026-10-21

Summary

VideoMaMa++ (arXiv 2610.26857) is a mask-guided diffusion framework for video matting that addresses temporal flickering at chunk boundaries. Diffusion-based matting benefits from video diffusion priors for zero-shot generalization from limited synthetic data, but heavy compute restricts inference to short chunks of 6-24 frames, and independent per-chunk matting causes flickering edges and discontinuous transparency. The proposed framework represents RGB and mask conditioning as separate token streams interacting through full spatio-temporal self-attention, and introduces learnable preserve tokens and refine tokens that enable sliding-window inference: previously generated matte results are propagated across chunks for local temporal consistency, while globally shared reference anchors prevent error accumulation over long sequences. Experiments on standard benchmarks and real long videos show superior detail accuracy and temporal consistency compared to prior diffusion-based matting methods. Authors: Sangbeom Lim, Seoung Wug Oh, Heeji Yoon, Seungryong Kim, Joon-Young Lee.

Paper Overview

Field: Computer Vision Authors: Sangbeom Lim, Seoung Wug Oh, Heeji Yoon, Seungryong Kim, Joon-Young Lee Published: 2026-10-19 arXiv: 2610.26857

Abstract

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 chunk of just 6 to 24 frames, so even ordinary videos span multiple chunks. Because each chunk is matted independently from coarse mask guidance, existing methods fail to maintain consistency across chunk boundaries, producing flickering edges and discontinuous transparency.

The authors address this with a mask-guided diffusion-matting framework built on two ideas:

1. Dual token streams: RGB and mask conditioning are represented as separate token streams that interact through full spatio-temporal self-attention, allowing the model to exploit both modalities effectively. 2. Preserve and refine tokens: Novel learnable preserve tokens and refine tokens enable sliding-window inference. Already-generated matte results are propagated across chunks to maintain local temporal consistency, complemented by globally shared reference anchors that prevent error accumulation over long sequences.

Results

On standard benchmarks and real-world long videos, the proposed framework outperforms previous diffusion-based matting methods in both detail accuracy and temporal consistency.