Submitted by myownskyW7 5 G^2RPO: Granular GRPO for Precise Reward in Flow Models IXCLab@Shanghai AI Lab 32 2