Problem
Research questions and friction points this paper is trying to address.
image editing
spatial credit
fine-grained reward
reinforcement learning
region-aware optimization
SpatialFlow-GRPO: Where Spatial Credit Drives Image Editing
🤖 AI Summary
Existing reinforcement learning approaches for image editing rely on global rewards, which hinder fine-grained spatial optimization. This work proposes SpatialFlow-GRPO, a novel framework that, for the first time, relaxes the assumption of reward spatial uniformity by introducing a spatially fine-grained reward mechanism. It transforms region-aware rewards into semantic region-level optimization signals and aligns regional advantages with their corresponding latent positions during policy gradient updates. To support this approach, we introduce the SFReward region-aware reward model along with the SFReward-14K dataset and propose MultiEditBench, a new benchmark for multi-region editing evaluation. Experiments demonstrate that our method significantly outperforms Flow-GRPO on both OmniGen2 and FLUX.2-klein-4B, achieving superior editing quality across GEdit-Bench, ImgEdit-Bench, and MultiEditBench.
📝 Abstract
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 is this spatial uniformity assumption: a whole-image reward cannot distinguish how different spatial regions contribute to image quality. To address this issue, we propose SpatialFlow-GRPO, a training framework that introduces spatially fine-grained reward feedback. The framework converts region-aware rewards into semantic-region-level optimization signals and aligns region advantages with the corresponding latent positions during policy updates. We also train a region-aware reward model, SFReward, construct SFReward-14K with region-annotated editing samples, and introduce MultiEditBench to evaluate multi-region editing ability. On OmniGen2 and FLUX.2-klein-4B, SpatialFlow-GRPO outperforms Flow-GRPO on GEdit-Bench, ImgEdit-Bench, and MultiEditBench. The results show that SpatialFlow-GRPO converts local feedback into spatially aligned update signals and improves editing quality.
Innovation
Methods, ideas, or system contributions that make the work stand out.
spatially fine-grained reward
region-aware optimization
SpatialFlow-GRPO
reinforcement learning for image editing
multi-region editing
🔎 Similar Papers
No similar papers found.
💼 Related Jobs
