Better, Stronger, Faster, and Broader: Structured All-Mask Prediction for MLLM-Based Segmentation

This paper introduces Structured All-Mask Prediction for multimodal large language model segmentation. Its STAMPlus model separates autoregressive dialogue from non-autoregressive mask prediction, generates a target list with explicit IDs and optional boxes, and jointly predicts multiple semantic or instance targets in a shared multiclass mask space. The abstract reports state-of-the-art results across referring, reasoning, open-vocabulary semantic, instance-aware, and remote-sensing small-target segmentation while preserving multimodal instruction following. For 12-category latency, repeated STAMP inference takes 13.50 seconds versus 5.16 seconds with STAMPlus. These claims are based on the supplied abstract and require verification against the full paper and benchmarks.