AI Paper Atlas

Research map / Video / Video generation

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

Sangbeom Lim, Seoung Wug Oh, Heeji Yoon, Seungryong Kim, Joon-Young Lee

NeurIPS 2026 · Video › Video generation · approach: Efficient video diffusion models

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VideoMaMa++, a diffusion-based video matting method that uses preserve-and-refine tokens to keep temporal consistency across independently processed chunks of frames.

Problem

Diffusion-based video matting is limited to short chunks of 6-24 frames due to heavy compute, and chunks are matted independently, causing temporal inconsistency across chunk boundaries for ordinary-length videos.

Approach

Introduces preserve-and-refine tokens to carry information across independently matted video chunks so that the matte stays temporally consistent; the abstract only partially describes this mechanism.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add reference anchor to sliding-window inference (ablation)VideoMaMa++ · VideoMatte (1920x1080)Grad: 2.2398 -> 1.7887Improved
add preserve token on overlap frames (ablation)VideoMaMa++ · Chunk boundary overlap discrepancyMAD: 0.7128 -> 0.2583Improved
use token concatenation instead of channel concatenation of RGB, mask and noise (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 2.2253 -> 1.6924Improved
condition on SAM logits instead of binary masks (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.6942 -> 1.6272Mixed
increase denoising steps from 1 to 5 at inference (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.6680 -> 10.7571Hurt
use chunk size 6 instead of 3 (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 2.1206 -> 1.6942Improved
use chunk size 12 instead of 6 (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.6942 -> 1.7906Hurt
increase LoRA rank from 64 to 512 (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 2.4294 -> 1.6942Improved
use random or last frame of first chunk as anchor instead of first frame (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.69 -> 1.70No clear effect
refresh anchor every 3 chunks (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.69 -> 1.70No clear effect
corrupt the anchor frame (full occlusion) (ablation)VideoMaMa++ · V-HIM60 (Hard)MAD: 1.71 -> 1.76Hurt
repeat interactive refinement iterations with preserve and refine tokensVideoMaMa++ · V-HIM60 (Hard)MAD: 1.6924 -> 1.6634Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Builds on: VideoMaMa

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy