Abstract

Current text-to-image diffusion models still struggle to follow complex and compositional prompts, largely due to the absence of consistent signals throughout the denoising trajectory. We introduce OUROBOROS, a reinforcement-guided flow matching framework that unifies critic optimization and generative diffusion in latent space. Inspired by ControlNet-style conditioning, our method incorporates a critic network that evaluates intermediate denoising steps via cross attention, producing reinforcement-based rewards for each latent update. These reward signals are used to iteratively refine the flow trajectory, enabling robust alignment between text and image over long-horizon, sparse-reward scenarios. To stabilize representation transitions across steps, we employ a cross-domain module inspired by ControlNet and CycleGAN, ensuring dimension consistency and cyclic regularization between encoder–decoder flows. We expect that the framework achieves scalable, bi-directional optimization that dynamically corrects latent paths while maintaining generation diversity.