arxivcs.CV2026-06-25
SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing
Yankai Yang, Yancheng Long, Wei Chen, Xingyu Lu, Hongyang Wei, Bin Wen, et al.
Recent online reinforcement learning has substantially improved image editing quality. However, existing Flow-GRPO-style methods usually rely on a single whole-image reward, which makes fine-grained editing optimization difficult. We observe that a key obstacle in image editing i…