The results support on-policy self-distillation as an efficient and analyzable approach to diffusion post-training by converting image-level reward guidance into explicit and continually refreshed intermediate supervision, thereby opening a path toward more efficient and diagnosable alignment.
Abstract
Reinforcement learning can align diffusion models with human preferences and task-specific objectives, but endpoint rewards do not specify how an intermediate denoising prediction should change. We introduce DiffusionOPSD as an on-policy self-distillation framework that converts image-level reward guidance into explicit targets for clean-output predictions at sampled queries. At each outer iteration, a frozen behavior policy generates trajectories and supplies query states and anchors. Reward gradients construct bounded positive and negative targets around each anchor. The trainable policy fits these targets as detached supervision through finite fitting before an exponential moving average update refreshes the behavior policy. This setup lets us measure target construction and finite realization separately. Controlled same-query experiments show that larger target-construction gains do not necessarily translate into larger realized gains after a single fitting update. Across SD 3.5-M and the step-distilled Z-Image-Turbo, our approach achieves the best final held-out scores in 19 of 20 reward-matched settings across two backbones and ten evaluators. It outperforms the strongest competing method by up to 44.0% and reduces training GPU-hours relative to DiffusionNFT by 40% on SD 3.5-M and 63% on Z-Image-Turbo. These results support on-policy self-distillation as an efficient and analyzable approach to diffusion post-training by converting image-level reward guidance into explicit and continually refreshed intermediate supervision, thereby opening a path toward more efficient and diagnosable alignment.
dOPSD derives the teacher's privilege directly from the student's own denoising trajectory, evaluating masked positions using later, more-decoded steps of that same trajectory rather than an external label, so the teacher's advantage emerges from the model's own decoding process.
Diffusion large language models (dLLMs) are a promising alternative to autoregressive generation. However, reasoning-oriented post-training for dLLMs remains challenging. Supervised fine-tuning (SFT) for dLLMs requires dense but often off-policy masked states, while reinforcement learning (RL) relies on sparse rewards or value modeling. This paper proposes \textbf{trace-based on-policy distillation (TOPD)}, a teacher-supervised framework that transfers reasoning ability to a target dLLM without reward estimation. The key idea is to supervise a dLLM on its own denoising trajectory, focusing on the trace-aligned token decisions that form the final response. Specifically, TOPD samples on-policy diffusion trajectories from the target dLLM, obtains teacher token distributions from a teacher model on the corresponding partially denoised states, and updates the target dLLM with a token-level Reverse Kullback-Leibler (Reverse-KL) objective. This design preserves dense teacher supervision while aligning training with the model's own denoising states. On mathematical reasoning benchmarks, TOPD enables SDAR-4B-Chat to match the MATH500 accuracy of its RL-trained counterpart TraDo-4B-Instruct, with gains of +5.7 under static evaluation and +4.5 under dynamic evaluation. Compared with the RL-trained counterpart, TOPD achieves this with 4$\times$ fewer rollout rounds, corresponding to an estimated 96.0$\times$ to-accuracy model-compute speedup.
Haolin Ren, Ziyang Huang, Chenhao Yuan et al.· 0 citations
Aligning diffusion models with human preferences usually relies on a sparse terminal reward evaluated on the final generated samples, which creates a severe temporal credit-assignment problem across the denoising process. We propose Latent Reward Registers, a mechanism that estimates terminal preference directly from intermediate noisy latents. Learnable, position-free register tokens are appended as an auxiliary read path to a frozen Diffusion Transformer (DiT), extracting preference signals without altering the generator's hidden states or velocity field. The resulting dense, differentiable reward field spans the full denoising trajectory and supports two alignment strategies. For training, Reward-Gradient On-Policy Distillation (RG-OPD) converts this dense reward field into per-step targets at states visited by the current generator, replacing rollout-intensive policy gradients with direct on-policy distillation. For inference, Reward-Guided Sampling (RGS) steers trajectories with magnitude-matched reward-gradient corrections and no parameter updates. Empirically, at high noise levels (t=0.8) the registers reach the highest pairwise accuracy among the evaluated latent reward models. RG-OPD outperforms online reinforcement learning baselines while reducing GPU hours by up to 33x. RGS achieves significant reward improvement with a favorable reward-quality balance against training-free baselines. Code and weights are to be available at https://github.com/Guanys-dar/latent-reward-register
Yuanshen Guan, Zipeng Feng, Chengru Song et al.· 0 citations
Reinforcement learning (RL) post-training provides a direct way to align diffusion models with human preferences and task-specific rewards. However, current RL algorithms for diffusion models remain fragmented: reverse-trajectory methods rely on discretized likelihood ratios, whereas forward-matching methods train on reward-labeled noising versions of the rollout samples. This paper shows that these seemingly different losses arise from a single path-space principle. Starting from the regularized diffusion-RL objective, we use importance sampling between sampling SDEs to obtain an explicit policy-gradient estimator on trajectory space. The estimator contains the stochastic It\^o integral underlying Flow-GRPO-type updates; we derive an equivalent variance-reduced value-gradient form that recovers the forward-matching structure of AWM and DiffusionNFT. This identifies the empirical gap between these method families as a variance-reduction effect rather than a difference in RL principle. The derivation yields a unified design space organized by value-gradient estimation, weight functions, and sampling choices. Within this space, we propose a multi-sample KDE value-gradient estimator that reuses rollout groups, together with scale-bounded weight families that retain stable existing recipes while excluding singular ones. Experiments on SD3.5-M and Qwen-Image models validate the variance-reduction explanation and show that the resulting recipe improves over prior diffusion-RL baselines.
Yixian Xu, Yuanrui Zhang, Shengjie Luo et al.· 1 citation
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. Reinforcement learning (RL) for preference alignment in diffusion models is promising but limited by reward sparsity. Since a single reward cannot support optimization, existing RL methods usually backpropagate the final reward to all previous steps. However, denoising is stage-wise, with distinct semantics and controllability. Repeating the final reward across all steps creates a temporal objective mismatch, encouraging reward shortcuts that lead to reward hacking. At the same time, due to reward backfilling, each time step receives the same reward, making it impossible to distinguish between actions, thereby weakening the optimization process. To resolve this issue, we propose Stage-Guided Per-Step Optimization (SGPO) for diffusion models, which jointly leverages signal-to-noise ratio and semantic changes to identify generation stages and adaptively assign stage-specific objectives. Early denoising is chaotic and far from the final reward, resulting in weak reward-behavior correlation. This stage should prioritize exiting the chaotic state. In the mid stage, the latent transitions to a stable structure, where the final reward better corresponds to generative behavior. Therefore, this stage optimizes the final reward while exploring diversity to avoid early convergence to a single mode. In the late stage, the latent's core structure is largely fixed, and preference optimization mainly amplifies local details, risking overfitting. Therefore, stable convergence is preferred to avoid quality degradation. Results from 16 comparative experiments validate SGPO. Our method achieves 26.7% average gains in generative quality and 36.7% higher convergence speed.
Renye Yan, Jikang Cheng, You Wu et al.· 0 citations
Reward fine-tuning is becoming an important tool for adapting diffusion models to human preferences and task-specific objectives, but existing methods largely inherit policy-gradient machinery from large language models. Unlike autoregressive models, diffusion models do not provide tractable likelihoods for generated samples. As a result, current approaches either construct trajectory likelihoods from stochastic denoising transitions or approximate endpoint likelihoods with evidence lower bound, introducing additional computation and algorithmic complexity. We demonstrate that this likelihood-based machinery is not necessary for effective diffusion reward fine-tuning. We propose reward-based velocity matching (RVM), a simple trajectory-free update that acts directly on the velocity field. RVM reinforces directions associated with high-reward generations, suppresses those with low reward, and involves an optional anchor term controlling drift from a reference velocity. Notably, it provides a general framework that recovers recent fine-tuning methods, including RAM and DiffusionNFT, as special cases. Across various large-scale diffusion models reward fine-tuning tasks, RVM is competitive with or outperforms trajectory-based policy-gradient methods under substantially reduced training cost. We further find that, once the velocity update is simplified, the particular loss variant matters less than reward and anchor design. For video generation, standard preference rewards can favor visually clean but nearly static outputs; introducing a new dynamic-tracking reward that substantially improve motions while improving overall VBench performance. These results suggest that scalable reward fine-tuning for diffusion models is better posed in the native velocity representation than as likelihood-based policy optimization.
Jaemoo Choi, Wei Guo, Yuchen Zhu et al.· 0 citations