The paper introduces DiffusionOPSD, an on‑policy self‑distillation framework that transforms image‑level reinforcement learning rewards into explicit targets for intermediate denoising predictions in diffusion models. By generating trajectories with a frozen behavior policy and constructing bounded positive and negative targets around query states, the method trains a policy to fit these targets before updating the behavior policy via an exponential moving average. Experiments on SD 3.5‑M and Z‑Image‑Turbo show that DiffusionOPSD achieves the best held‑out scores in 19 of 20 reward‑matched settings, outperforms the strongest competitor by up to 44 % and cuts GPU‑hour usage by 40–63 % compared to DiffusionNFT.
By Wei Zhou, Xiongwei Zhu, Lingdong Kong, Bo Chen, Lei Zhang, Yongyuan Liang, Xiaoxia Hou, Ye Tian, Xian Sun, Yingshuo Wang, Linfeng Li, Shengqiong Wu, Leigang Qu, Feng Li, Wei Liu, Julian McAuley, Tat-Seng Chua
arXiv:2602. 07339v2 Announce Type: replace Abstract: Diffusion-based trajectory planners can model multi-modal driving behavior, but their iterative denoising process introduces a latency bottleneck for real-time closed-loop deployment.
By Ruturaj Reddy, Hrishav Bakul Barua, Junn Yong Loo, Thanh Thi Nguyen, Ganesh Krishnasamy
arXiv:2608. 03929v1 Announce Type: new Abstract: Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, presenting a severe temporal credit-assignment challenge across the multi-step denoising process.
By Yuanshen Guan, Zipeng Feng, Zhiwei Xiong, Peiqin Sun
DiffusionART introduces a new reinforcement learning framework for diffusion models that aggregates prompt rewards into a shared direction, eliminating consistency forces inherent in prior methods. It employs a trust‑region‑masked second update for drift‑adaptive restoration, removing the need for EMA tuning and differentiable reward models. Experiments on image and video generation show that DiffusionART trains more efficiently and stably, achieving higher mean rewards with significantly fewer GPU hours compared to existing approaches.
By Zhenglin Wan, Jiayu Chen, Zhengfeiyang Zhang, Pengfei Zhou, Wangbo Zhao, Fangyuan Kong, Pengfei Wan, Kun Gai, Yang You
Diffusion models have strong generative capabilities. However, their maximum likelihood training objective only focuses on reconstructing the data distribution, making it difficult to align with specific preferences.
arXiv:2608.23664v1 Announce Type: cross
Abstract: Reward fine-tuning is becoming an important tool for adapting diffusion models to human preferences and task-specific objectives, but existing method...
By Jaemoo Choi, Wei Guo, Yuchen Zhu, Arash Vahdat, Molei Tao, Julius Berner, Yongxin Chen
Efficient text-to-image generation requires both reinforcement-learning (RL)-based reward alignment and few-step distillation, yet these procedures are typically performed sequentially, increasing training cost and risking the loss of reward gains during compression. We instead take an RL-native perspective: diffusion RL already generates reward-scored finite-step trajectories, whose intermediate states provide a natural source of distillation supervision rather than a disposable byproduct of sampling.
The paper introduces Diffusion-Augmented Markov Decision Processes (DA‑MDPs), a framework that extends Maximum Entropy Reinforcement Learning to diffusion-based policies. DA‑MDPs treat each reverse‑diffusion step as an RL decision, deriving a tractable reverse‑KL bound that decomposes across denoising transitions and yields diffusion‑augmented soft rewards, value functions, and policy objectives. The authors implement this framework with PPO, REPPO, and a maximum‑entropy WPO variant, showing improved continuous‑control performance, higher success rates on manipulation tasks, and memory‑efficient training with action chunking.
By Sebastian Sanokowski, Kaustubh Patil, Majid Khadiv
CAST introduces a reinforcement‑learning fine‑tuning framework for diffusion models that addresses three key limitations: it automatically selects the denoising window based on each model’s trajectory, decomposes prompts into verifiable semantic atoms via Causal Scene Graphs, and applies atom‑level rewards spatially weighted in the policy objective. The method is applied to FLUX.2‑dev and Qwen‑Image‑2512, yielding up to 3.07× improvement on the hardest GenEval 2 prompts compared with Flow‑GRPO while also enhancing overall generation quality.
By Shu Yu, Chaochao Lu
Routing Policy Optimization (RPO) introduces stochastic exploration within a diffusion visual generator by treating attention routing as a stochastic action. Instead of perturbing the entire reverse denoising trajectory, RPO samples key attention blocks at selected denoising levels, perturbs the transition velocity, and evaluates the perturbation with a score‑function objective. A two‑rollout estimator provides an advantage signal, and after training the model restores dense attention, yielding higher generation quality and nearly four times the training efficiency compared to Flow‑GRPO‑style methods.
arXiv:2607. 07693v1 Announce Type: cross Abstract: Reinforcement learning from human feedback (RLHF) has emerged as a powerful paradigm for aligning generative models with human preferences.
By Eric Zhu, Abhinav Shrivastava, Soumik Mukhopadhyay
DiffusionOPD introduces a multi-task training framework for diffusion models that leverages Online Policy Distillation (OPD). The method trains task-specific teachers separately and then distills their knowledge into a single student model using the student's own rollout trajectories, thereby separating exploration from integration. The authors extend OPD from discrete tokens to continuous-state Markov processes, deriving a closed-form per-step KL objective that unifies stochastic SDE and deterministic ODE refinement, and show that this analytic gradient yields lower variance and better generality than PPO-style gradients. Experiments demonstrate that DiffusionOPD outperforms both multi-reward RL and cascade RL baselines in training efficiency and final performance, achieving state-of-the-art results across all evaluated benchmarks.
By Quanhao Li, Junqiu Yu, Kaixun Jiang, Yujie Wei, Zhen Xing, Pandeng Li, Ruihang Chu, Shiwei Zhang, Yu Liu, Zuxuan Wu