# On-Policy Self-Distillation in Diffusion Models

arXiv: https://arxiv.org/abs/2608.24646 · HTML: https://mlsanity.com/paper/2608.24646/

## Key facts

- Cited by 7 papers in the map (data through 14 Oct 2026).
- Placed in Diffusion and flow policies for robotics, one of 14 directions in Diffusion models.
- First posted on arXiv 25 Aug 2026.

## Placed in

- Diffusion models / [Diffusion and flow policies for robotics](https://mlsanity.com/diffusion/diffusion-and-flow-policies-for-robotics/)
- Deep reinforcement learning methods / [RL with diffusion and flow models](https://mlsanity.com/deeprl/rl-with-diffusion-and-flow-models/)

## Built on by (most cited first, 7 of 7)

- [SenseNova-U1.5: Towards Native Unified Visual Intelligence](https://mlsanity.com/paper/2609.11929/) (2026)
- [OracleZoom: On-Policy Self-Distillation Inspired Reference-Constrained Recursive Image Super Resolution](https://mlsanity.com/paper/2609.06490/) (2026)
- [On-Policy Self-Distillation for Multi-Turn Image Editing](https://mlsanity.com/paper/2609.35611/) (2026)
- [UniEvo-VL: An On-policy Self-Distillation Training Recipe for Multimodal Model Self-improvement](https://mlsanity.com/paper/2609.38721/) (2026)
- [MEND: RL For Flow Models via Proximal Velocity Matching](https://mlsanity.com/paper/2610.05954/) (2026)
- [A Lightweight Velocity Correction Framework for Improving Instruction Following in Few-Step Text-to-Image Generation](https://mlsanity.com/paper/2610.12899/) (2026)
- [VITAL: From Visual Motion to Precise Action](https://mlsanity.com/paper/2610.20604/) (2026)

## Abstract (arXiv)

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.