{"paper":{"title":"Sa2VA: Marrying SAM2 with LLaVA for Dense Grounded Understanding of Images and Videos","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Sa2VA unifies segmentation and language models for referring tasks on both images and videos using minimal instruction tuning.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Haobo Yuan, Jiashi Feng, Lu Qi, Ming-Hsuan Yang, Shilin Xu, Shunping Ji, Tao Zhang, Xiangtai Li, Yueyi Sun, Yunhai Tong, Zilong Huang","submitted_at":"2025-01-07T18:58:54Z","abstract_excerpt":"This work presents Sa2VA, the first comprehensive, unified model for dense grounded understanding of both images and videos. Unlike existing multi-modal large language models, which are often limited to specific modalities and tasks, Sa2VA supports a wide range of image and video tasks, including referring segmentation and conversation, with minimal one-shot instruction tuning. Sa2VA combines SAM-2, a foundation video segmentation model, with MLLM, the advanced vision-language model, and unifies text, image, and video into a shared LLM token space. Using the LLM, Sa2VA generates instruction to"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Sa2VA is the first comprehensive, unified model for dense grounded understanding of both images and videos that supports referring segmentation and conversation with minimal one-shot instruction tuning.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the LLM-generated instruction tokens can reliably guide SAM-2 to produce precise masks across complex video scenes without task-specific architectural changes or heavy fine-tuning.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Sa2VA unifies SAM-2 segmentation with MLLM reasoning into a single model for referring segmentation and conversation on images and videos, supported by a new 72k-expression Ref-SAV dataset.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Sa2VA unifies segmentation and language models for referring tasks on both images and videos using minimal instruction tuning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e2e0c6f450bbd5a684e1cce0175ca33a6167b5f731f0c3ab33d5f25dd01d6f71"},"source":{"id":"2501.04001","kind":"arxiv","version":3},"verdict":{"id":"96ea7224-7912-48dc-ba16-361d3a3b6fc6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T11:35:57.949381Z","strongest_claim":"Sa2VA is the first comprehensive, unified model for dense grounded understanding of both images and videos that supports referring segmentation and conversation with minimal one-shot instruction tuning.","one_line_summary":"Sa2VA unifies SAM-2 segmentation with MLLM reasoning into a single model for referring segmentation and conversation on images and videos, supported by a new 72k-expression Ref-SAV dataset.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the LLM-generated instruction tokens can reliably guide SAM-2 to produce precise masks across complex video scenes without task-specific architectural changes or heavy fine-tuning.","pith_extraction_headline":"Sa2VA unifies segmentation and language models for referring tasks on both images and videos using minimal instruction tuning."},"references":{"count":125,"sample":[{"doi":"","year":2015,"title":"Vqa: Visual question an- swering","work_id":"4f7b2ba9-84f5-40f4-8098-9380e897823d","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond","work_id":"cbc2bb21-b6bb-46c0-80bf-107e195ffe10","ref_index":2,"cited_arxiv_id":"2308.12966","is_internal_anchor":true},{"doi":"","year":2025,"title":"Qwen2.5-VL Technical Report","work_id":"69dffacb-bfe8-442d-be86-48624c60426f","ref_index":3,"cited_arxiv_id":"2502.13923","is_internal_anchor":true},{"doi":"","year":2024,"title":"One token to seg them all: Language instructed reasoning segmentation in videos","work_id":"530daa38-715a-4558-b3a3-a6815a7b7604","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Language models are few-shot learners","work_id":"0640f028-a03b-4b3e-b61b-59f4dc45f798","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":125,"snapshot_sha256":"e92b0993659e5cb959521db9c098eb29a88fa395e46bdf26e2ead24e0756b1c3","internal_anchors":21},"formal_canon":{"evidence_count":2,"snapshot_sha256":"c5a0eaf80ac15fff66dc43e957e0b934c253c88c53d0a75462d6131576c315cb"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}