Sa2VA

Why we included this project

Sa2VA combines the SAM-2 segmentation model with a multimodal LLM, so a single model can discuss images and videos and return pixel-level masks. That covers referring segmentation, grounded conversation, visual prompting, and chat. The repo also contains several companion projects: VRT adds object-level grounded reasoning with its own benchmark and training data, SAMTok turns masks into tokens any MLLM can use, and SaSaSa2VA, a segmentation-augmented variant, won first place in the ICCV 2025 LSVOS challenge. You can pick between several backbones, including InternVL2.5/3 and Qwen2.5/3-VL, download pretrained weights from Hugging Face, and run a Gradio demo before integrating. If you're building annotation tools, video understanding, or promptable segmentation systems, this is a solid foundation.