Reinforce Your Layout: Online Reward-Guided Diffusion for Layout-to-lmage Generation

In this work, we tackle the layout-to-image generation task by proposing a novel online reinforcement learning (RL) framework that directly optimizes diffusion models to achieve consistency between images and layouts. We introduce RLLay, a method that overcomes a major limitation that lies in existing methods with their reliance on indirect side guidance—rather than direct supervision on layout alignment—which constrains these models' ability to accurately position and scale image content. Given a prompt, our approach generates multiple candidate images and ranks them using a reward model based on Intersection-over-Union (IoU) to quantify alignment between predicted and target layouts. To effectively utilize this ranking signal, we introduce a pairwise preference-based optimization strategy that fine-tunes the diffusion model by maximizing the likelihood of higher-ranked samples relative to lower-ranked ones (hard-negatives). Experimental results show that our RL-based fine-tuning significantly improves both spatial layout fidelity and text-image alignment, establishing a promising direction for more controlled and layout-aware image generation.