AIAny
Icon for item

On-Policy Self-Distillation in Diffusion Models

Converts image-level rewards into explicit intermediate targets for diffusion-model denoising via an on-policy self-distillation loop. Constructs bounded positive/negative targets around anchors from reward gradients, fits those targets with finite updates, and refreshes a behavior policy by EMA—improving aligned performance across backbones while reducing GPU hours.

Introduction

Most reward-based post-training of diffusion models leaves intermediate denoising steps unspecified: an image-level reward tells you which final samples are better, but not how each denoising prediction should change. The core insight of this work is that converting image-level reward signals into explicit, bounded per-query targets—and refreshing those targets on the student’s own rollouts—makes alignment both more effective and more analyzable.

Key Findings
  • Converting image-level reward gradients into bounded positive and negative targets around anchors yields dense, per-query supervision that the model can directly fit.
  • Separating target construction from finite realization exposes when larger constructed gains fail to translate after a single fitting update, enabling diagnostic insight into training dynamics.
  • Across two backbones (SD 3.5-M and a step-distilled Z-Image-Turbo) and ten evaluators, the method achieves the top held-out scores in 19 of 20 reward-matched settings, outperforms the strongest competing method by up to 44.0%, and reduces GPU-hours versus DiffusionNFT by ~40% (SD 3.5-M) and ~63% (Z-Image-Turbo).
Method overview
  • Use a frozen behavior policy to generate on-policy denoising trajectories and collect query states and anchors.
  • Build bounded positive/negative targets at each query from reward gradients around the anchor prediction.
  • Train the student to fit these detached targets via a finite number of fitting updates; then refresh the behavior policy with an exponential moving average (EMA) of the student.
  • Repeat: construction of targets and finite fitting are alternated so that target quality and realizable gains can be measured independently.
Who it's for and trade-offs

Great fit if you want diagnosable, compute-efficient post-training alignment of image-generation diffusion models and can run on-policy rollouts and short fitting cycles. Look elsewhere if you cannot afford on-policy sampling at scale, have no reward signal to shape targets, or require purely off-policy supervised fine-tuning without changing the rollout distribution. The approach introduces hyperparameters for target bounds and fitting steps that must be tuned; gains depend on the interplay between target construction and the student’s capacity to realize those targets.

Information

  • Websitearxiv.org
  • AuthorsWei Zhou, Xiongwei Zhu, Lingdong Kong, Bo Chen, Lei Zhang, Yongyuan Liang, Xiaoxia Hou, Ye Tian, Xian Sun, Yingshuo Wang
  • Published date2026/08/25

More Items

Proposes Recuris, a recursive Experiential-Working Memory architecture that separates Working Memory (task progress) from Experiential Memory (skills) and uses a Meta-Agent to validation-gate localized skill updates, enabling bounded recursive skill evolution for long-horizon agents.

Generates unified embeddings for text, images, video, visual documents and interleaved multimodal inputs with configurable output dimensions and Matryoshka truncation to trade accuracy for cost. Model weights and code are released under Apache-2.0; the 9B variant scores 80.6 on MMEB-v2.

Treats human annotations as oracle rollouts and separates them from on-policy baselines to improve reinforcement learning for video multimodal LLMs. Key features include a decoupled advantage estimator, sign-balanced pruning, and scalable gains across model sizes and data budgets.