Seg-Zero

Why we included this project

Seg-Zero is the official code for a research line that gets vision-language models to segment images by reasoning about what to look for first. Training is pure reinforcement learning: IoU and L1 rewards push the model to emit a reasoning chain before the final mask, with no supervised reasoning data in the loop. If you work on referring expression segmentation or visual grounding, the released 7B checkpoints and the RefCOCOg-derived training data give you a concrete starting point. The code builds on EasyR1 and veRL, supports Qwen2-VL and Qwen2.5-VL, and includes inference and evaluation scripts plus GRPO training for multi-GPU setups. It is research software rather than a turnkey service, so expect to adapt the pipeline to your own data and hardware.