SAMTok: Representing Any Mask with Two Words
Paper • 2601.16093 • Published • 44
SAMTok
[CVPR 2026 Highlight] A unified mask-token interface for MLLMs.
-
-
SAMTok
🎨24SAMTok provides a unified mask-token interface for MLLMs.
-
zhouyik/Qwen3-VL-8B-SAMTok
Image-Text-to-Text • 9B • Updated • 69 • 1 -
zhouyik/Qwen3-VL-4B-SAMTok
Image-Text-to-Text • 5B • Updated • 71 • 1
zhouyik/Qwen2.5-VL-7B-SAMTok-co
Image-Text-to-Text • 8B • Updated • 29Note Experimental checkpoint co-trained on mask generation and mask understanding tasks.
zhouyik/Qwen2.5-VL-3B-SAMTok-co
Image-Text-to-Text • 4B • Updated • 40Note Experimental checkpoint co-trained on mask generation and mask understanding tasks.
zhouyik/Qwen3-VL-4B-SAMTok-co
Image-Text-to-Text • 4B • Updated • 31Note Experimental checkpoint co-trained on mask generation and mask understanding tasks.
zhouyik/PLM-1B-SAMTok-co
Image-Text-to-Text • 2B • Updated • 26Note Experimental checkpoint co-trained on mask generation and mask understanding tasks.
zhouyik/Qwen3-VL-4B-SAMTok-dam
Image-Text-to-Text • 4B • Updated • 27Note Dedicated mask-understanding checkpoint for region-conditioned text generation.
zhouyik/Qwen2.5-VL-3B-SAMTok-gcg-rl
Image-Text-to-Text • 4B • Updated • 27Note Task-specialized checkpoint post-trained with GRPO on GCG using a textual mask answer-matching reward.
zhouyik/Qwen2.5-VL-3B-SAMTok-gres-rl
Image-Text-to-Text • 4B • Updated • 34 • 2Note Task-specialized checkpoint post-trained with GRPO on GRES using a textual mask answer-matching reward and a thinking-answer format.
zhouyik/Qwen2.5-VL-3B-SAMTok-gcg-ft
Image-Text-to-Text • 4B • Updated • 28Note Task-specialized checkpoint further supervised fine-tuned on grounded conversation generation (GCG).
zhouyik/Qwen2.5-VL-7B-SAMTok-gcg-ft
Image-Text-to-Text • 8B • Updated • 22Note Task-specialized checkpoint further supervised fine-tuned on grounded conversation generation (GCG).
zhouyik/Qwen2.5-VL-3B-SAMTok-gres-ft
Image-Text-to-Text • 4B • Updated • 23Note Task-specialized checkpoint further supervised fine-tuned on generalized referring expression segmentation (GRES).
zhouyik/Qwen2.5-VL-7B-SAMTok-gres-ft
Image-Text-to-Text • 8B • Updated • 28Note Task-specialized checkpoint further supervised fine-tuned on generalized referring expression segmentation (GRES).