ENHANCING DIFFUSION MODELS WITH REINFORCEMENT LEARNING

Summary from the TLDR AI newsletter

This paper investigates the use of reinforcement learning to optimize diffusion models for downstream objectives like image quality. The proposed approach, Denoising Diffusion Policy Optimization (DDPO), proves effective in adapting text-to-image models to challenging objectives and improving prompt-image alignment without additional data or human annotation.