Researchers develop DecAF for training-free video reasoning segmentation

Researchers have developed Decomposed Attention Fusion (DecAF), a novel method for video reasoning segmentation that operates without requiring model retraining. DecAF refines attention maps generated by multimodal large language models (MLLMs) by contrasting object and background activations and fusing complementary frame-level attention. This approach allows for the direct conversion of attention maps into segmentation masks, achieving performance comparable to training-based methods on video object segmentation benchmarks. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Enables training-free video segmentation by refining MLLM attention maps, potentially simplifying deployment for video analysis tasks.

RANK_REASON This is a research paper detailing a new method for video reasoning segmentation.

Read on arXiv cs.CV →

• Decomposed Attention Fusion
• DecAF
• MLLMs
• Jeongseok Hyun
• SAM
• arXiv

• paper
• model release