Enhancing Diffusion Models with RL and Adversarial Rewards
Overview
A research project exploring the intersection of reinforcement learning and diffusion models for image generation.
Key Contributions
- Novel MDP formulation of the reverse diffusion process with adversarial discriminators as reward signals
- 21.7% FID reduction compared to baseline diffusion models
- Plug-and-play design — can be applied to existing pretrained diffusion models without retraining from scratch