Better, Stronger, Faster, and Broader: Structured All-Mask Prediction for MLLM-Based Segmentation
STAMPlus predicts multiple segmentation targets in one non-autoregressive pass while keeping dialogue behavior intact.
The paper frames MLLM segmentation as a tradeoff between mask quality, instruction-following ability, and speed. Its STAMPlus method generates target IDs, optionally uses boxes, and binds those IDs to a shared multi-class mask space. The authors say one checkpoint covers referring, reasoning, open-vocabulary semantic, instance-aware, and remote-sensing small-target segmentation. They report 12-category latency falling from 13.50 seconds with repeated STAMP inference to 5.16 seconds.
HF Daily Papers' note
The paper frames MLLM segmentation as a tradeoff between mask quality, instruction-following ability, and speed. Its STAMPlus method generates target IDs, optionally uses boxes, and binds those IDs to a shared multi-class mask space. The authors say one checkpoint covers referring, reasoning, open-vocabulary semantic, instance-aware, and remote-sensing small-target segmentation. They report 12-category latency falling from 13.50 seconds with repeated STAMP inference to 5.16 seconds.
HF Daily Papers' note
score 5