SpatialReward: Bridging the Perception Gap in Online RL for Image Editing via Explicit Spatial Reasoning

Online Reinforcement Learning (RL) offers a promising avenue for complex image editing but is currently constrained by the scarcity of reliable and fine-grained reward signals. Existing evaluators frequently struggle with a critical perception gap we term "Attention Collapse," where models neglect cross-image comparisons and fail to capture fine-grained details, resulting in inaccurate perception and miscalibrated scores. To address these limitations, we propose SpatialReward, a reward model that enforces precise verification via explicit spatial reasoning. By anchoring reasoning to predicted edit regions, SpatialReward grounds semantic judgments in pixel-level evidence, significantly enhancing evaluative accuracy. Trained on a curated 260k spatial-aware dataset, our model achieves state-of-the-art performance on MMRB2 and EditReward-Bench, and outperforms proprietary evaluators on our proposed MultiEditReward-Bench. Furthermore, SpatialReward serves as a robust signal in online RL, boosting OmniGen2 by +0.90 on GEdit-Bench—surpassing the leading discriminative model and doubling the gain of GPT-4.1 (+0.45). These results demonstrate that spatial reasoning is essential for unlocking effective alignment in image editing.

Image editing AI systems can follow many written instructions, such as adding an object, changing a color, or preserving part of an image. However, it is still difficult to automatically judge whether an edited image truly follows the instruction, especially when the required changes involve several objects or precise locations. Existing scoring methods often look at the image too broadly and may miss small but important mistakes, which can give misleading feedback when training better editing models. In this paper, we introduce SpatialReward, a scoring model that checks image edits with explicit attention to where changes should happen. Instead of only asking whether the whole image looks good, SpatialReward reasons about specific regions and compares the original and edited images more carefully. We also build a challenging benchmark, MER-Bench, to test models on complex editing tasks with multiple requirements. Our results show that SpatialReward agrees better with human judgments and provides more reliable feedback for improving image editing systems. This suggests that spatial reasoning is important for making image editing AI more accurate and trustworthy.