The paper investigates how reinforcement learning can be effectively applied to diffusion models for visual tasks, focusing on the role of likelihood estimation. By systematically separating policy‑gradient objectives, likelihood estimators, and rollout sampling schemes, the authors find that using an evidence lower bound (ELBO) based likelihood estimator computed from the final generated sample is the key factor for stable and efficient RL optimization, outweighing the choice of loss function. Experiments on SD 3.5 Medium across multiple reward benchmarks confirm that this approach improves GenEval scores from 0.24 to 0.95 in 90 GPU hours, outperforming existing methods such as FlowGRPO and the current state‑of‑the‑art without reward hacking.
By Jaemoo Choi, Yuchen Zhu, Wei Guo, Petr Molodyk, Bo Yuan, Jinbin Bai, Yi Xin, Molei Tao, Yongxin Chen
arXiv:2609.37227v1 Announce Type: new
Abstract: Inference-time steering adapts pretrained diffusion and flow-based models to new tasks, e.g., to generate samples from a conditional distribution or sa...
By Adhithyan Kalaivanan, Zheng Zhao, Jens Sj\"olund, Fredrik Lindsten
arXiv:2609.31882v2 Announce Type: replace-cross
Abstract: Reward-based diffusion fine-tuning faces practical challenges when desirable outcomes are rare or conditioning corrections are costly to esti...
By Zhengyi Guo, Jiayuan Sheng, Wenpin Tang, David D. Yao
Diffusion models are increasingly used as controllable samplers, whose generations can be steered at inference time according to a chosen reward function. While such rewards are typically defined on individual samples, for many applications it is desirable to steer according to distribution-level rewards, for example to calibrate with population-level information or to encourage diversity.
arXiv:2608. 08770v1 Announce Type: cross Abstract: Diffusion models are increasingly used as controllable samplers, whose generations can be steered at inference time according to a chosen reward function.
By Samuel Howard, Nikolas N\"usken
The paper introduces a data‑free learning objective called relative trajectory balance for training diffusion models to sample from a posterior defined by a diffusion prior and an arbitrary black‑box constraint or likelihood. It proves asymptotic correctness of this objective and demonstrates its use across vision, language, and multimodal tasks, including classifier guidance, language infilling, and text‑to‑image generation. Additionally, the method is applied to continuous control with a score‑based behavior prior, achieving state‑of‑the‑art results in offline reinforcement learning.
By Siddarth Venkatraman, Moksh Jain, Luca Scimeca, Minsu Kim, Marcin Sendera, Mohsin Hasan, Luke Rowe, Sarthak Mittal, Pablo Lemos, Emmanuel Bengio, Alexandre Adam, Jarrid Rector-Brooks, Yoshua Bengio, Glen Berseth, Esmeralda S. Whitammer
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
The paper introduces GRAS, a method that improves training‑free reward alignment for discrete diffusion models by reducing variance in guided proposals and adapting the resampling temperature during search. It achieves this without adding denoiser cost, using Rao‑Blackwellized estimates for differentiable rewards and a leave‑one‑out baseline for non‑differentiable ones. Experiments on regulatory DNA and protein design show GRAS outperforms existing training‑free techniques and rivals reward‑fine‑tuned models.
By Kwanyoung Kim
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: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
arXiv:2602. 03211v2 Announce Type: replace-cross Abstract: Diffusion models have demonstrated strong generative performance; however, generated samples often fail to fully align with human intent.
By Yeongmin Kim, Donghyeok Shin, Byeonghu Na, Minsang Park, Richard Lee Kim, Il-Chul Moon
VGAS: Variance-Reduced Guidance and Adaptive Selection for Training-Free Reward Alignment in Discrete Diffusion proposes a new inference-time framework that improves steering of frozen masked discrete diffusion models. By reducing the variance of guidance estimates, applying reward tilting to clean-token logits, and adapting the selection temperature at each step, VGAS addresses three default choices in existing pipelines. Experiments on regulatory DNA, protein, and small-molecule benchmarks show that VGAS achieves the best training-free reward performance and matches or surpasses reward-fine-tuned generators.
By Kwanyoung Kim