DiffusionOPD

A Unified Perspective of On-Policy Distillation in Diffusion Models

   

Model Card

This Hugging Face repository hosts the released LoRA checkpoints for DiffusionOPD, an online policy distillation framework for multi-task diffusion alignment.

DiffusionOPD first trains task-specialized teacher models and then distills their capabilities into one unified student along the student's own rollout trajectories. The released checkpoints are intended to support reproducible academic research on diffusion alignment, reward optimization, and on-policy distillation.

Released Checkpoints

The repository contains LoRA checkpoints trained from stabilityai/stable-diffusion-3.5-medium:

Directory	Description
Student/lora	Unified DiffusionOPD student distilled from multiple task-specialized teachers.
AesTeacher/lora	Teacher optimized for aesthetic preference alignment.
GenEvalTeacher/lora	Teacher optimized for GenEval-style compositional generation.
OCRTeacher/lora	Teacher optimized for OCR/text rendering capability.

Base Model

These checkpoints are LoRA adapters for:

stabilityai/stable-diffusion-3.5-medium

Powered by Stability AI.

Usage

Please use the official DiffusionOPD codebase for loading the LoRA checkpoints and running evaluation:

git clone https://github.com/ali-vilab/DiffusionOPD.git
cd DiffusionOPD

conda create -n DiffusionOPD python=3.10.16
conda activate DiffusionOPD
pip install torch==2.6.0 torchvision==0.21.0 --index-url https://download.pytorch.org/whl/cu126
pip install -e .

bash scripts/single_node/eval.sh

The evaluation script supports datasets including geneval, ocr, pickscore, and drawbench.

Intended Use

This project is released for academic research only. It is intended for studying diffusion model alignment, reward-guided optimization, on-policy distillation, and evaluation of text-to-image generation systems.

Users are responsible for ensuring that their use of these checkpoints complies with all applicable laws, research ethics requirements, the Creative Commons Attribution license for this release, and the license terms of the underlying Stability AI model.

License

The DiffusionOPD released checkpoints and model card are provided under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Because these checkpoints are based on stabilityai/stable-diffusion-3.5-medium, use of the base model is also governed by the Stability AI Community License:

https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/blob/main/LICENSE.md

Required Stability AI attribution:

"This Stability AI Model is licensed under the Stability AI Community License, Copyright © Stability AI Ltd. All Rights Reserved"

Powered by Stability AI

Citation

If you find this project useful, please cite:

@article{li2026diffusionopd,
  title={DiffusionOPD: A Unified Perspective of On-Policy Distillation in Diffusion Models},
  author={Li, Quanhao and Yu, Junqiu and Jiang, Kaixun and Wei, Yujie and Xing, Zhen and Li, Pandeng and Chu, Ruihang and Zhang, Shiwei and Liu, Yu and Wu, Zuxuan},
  journal={arXiv preprint arXiv:2605.15055},
  year={2026}
}

Acknowledgements

We thank the Flow-GRPO, DiffusionNFT, and Stable Diffusion 3.5 Medium projects.