DiffusionART: Efficient, Stable and Scalable Reinforcement of Diffusion Models
DiffusionART addresses instability in online reinforcement learning for diffusion models by introducing a method that removes a hidden forward-reverse consistency force. It achieves this by aggregating prompt rewards into a shared direction and using a trust-region-masked second update for drift-adaptive restoration, eliminating the need for EMA tuning and per-task training. This results in more efficient and stable training across image and video generation tasks.