# bytedance/Sa2VA

Official Repo For Pixel-LLM Codebase: Sa2VA (T-PAMI-26), SAMTok (CVPR-26), VRT (Arxiv-25), SaSaSa2VA (1-st solution for LSVOS)

Repository: https://github.com/bytedance/Sa2VA
Canonical: https://ross.abutalabs.com/products/sa2va
Language: Python
License: Apache-2.0
License Family: permissive
Topics: computer-vision, mllm, large-language-models
Last push: 2026-08-04T23:56:43+00:00

## Health v2 (maintenance only)
Score: 70/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 96, release rhythm 52, longevity 43
- inputs: {"age_days": 604, "days_push": 29, "days_rel": 320, "gap_med": 1, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1666, forks 131 (observed 2026-08-28T04:05:19.111941+00:00)

## What it is
Sa2VA is a family of research models and codebases from ByteDance that combine SAM-2 with multimodal LLMs for pixel-level grounded understanding of images and videos. It includes the core Sa2VA model plus extensions like SAMTok, VRT, Pixel-SAIL, and SaSaSa2VA, with pretrained checkpoints on Hugging Face.

## Use cases
- referring image and video segmentation with a multimodal LLM
- grounded conversation about images and videos
- chat with an AI about specific pixels or regions in an image
- generate segmentation masks from natural language instructions
- visual prompting and region-level question answering
- evaluate grounded visual reasoning with VRT-Bench
- train a multimodal LLM that outputs segmentation masks
- video object segmentation benchmark submission

## When to choose
- you need unified referring segmentation and visual chat in one model
- you want pixel-level grounding on top of InternVL or Qwen-VL backbones
- you are doing research on grounded multimodal understanding
- you need pretrained checkpoints for image and video segmentation tasks

## When to avoid
- you need a production-ready, latency-optimized vision API
- you lack a GPU or cannot run large multimodal models
- you only need simple image classification or object detection
- you want a no-code GUI tool rather than a Python research codebase

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, computer-vision, deep-learning, nlp, image-processing, video-processing
- domain: computer-vision, large-language-models, machine-learning, artificial-intelligence
- platform: python
- tags: multimodal-llm, segmentation, sam2, grounded-understanding, referring-segmentation, visual-prompting, research-code, image-chat, video-chat, mask-tokens, gpu, linux

## Member repositories
- bytedance/Sa2VA (main) score 70

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:05:19.111941+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T03:42:57.832601+00:00, confidence not recorded.
  - readme: https://github.com/bytedance/Sa2VA (fetched 2026-08-28T04:05:19.111941+00:00, sha b80c36c49b30)
- Data as of 2026-08-30T08:39:29.467469+00:00.