English

Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion

Computer Vision and Pattern Recognition 2026-03-03 v1 Artificial Intelligence Machine Learning

Abstract

Text-to-image diffusion models often memorize training data, revealing a fundamental failure to generalize beyond the training set. Current mitigation strategies typically sacrifice image quality or prompt alignment to reduce memorization. To address this, we propose Reachability-Aware Diffusion Steering (RADS), an inference-time framework that prevents memorization while preserving generation fidelity. RADS models the diffusion denoising process as a dynamical system and applies concepts from reachability analysis to approximate the "backward reachable tube"--the set of intermediate states that inevitably evolve into memorized samples. We then formulate mitigation as a constrained reinforcement learning (RL) problem, where a policy learns to steer the trajectory away from memorization via minimal perturbations in the caption embedding space. Empirical evaluations show that RADS achieves a superior Pareto frontier between generation diversity (SSCD), quality (FID), and alignment (CLIP) compared to state-of-the-art baselines. Crucially, RADS provides robust mitigation without modifying the diffusion backbone, offering a plug-and-play solution for safe generation. Our website is available at: https://s-karnik.github.io/rads-memorization-project-page/.

Keywords

Cite

@article{arxiv.2603.00140,
  title  = {Steering Away from Memorization: Reachability-Constrained Reinforcement Learning for Text-to-Image Diffusion},
  author = {Sathwik Karnik and Juyeop Kim and Sanmi Koyejo and Jong-Seok Lee and Somil Bansal},
  journal= {arXiv preprint arXiv:2603.00140},
  year   = {2026}
}
R2 v1 2026-07-01T10:56:19.151Z