DMDR:
Distribution Matching Distillation Meets Reinforcement Learning

---

Images generated by Z-Image-Turbo.

---

Images generated by SD3.5 Large finetuned through DMDR with open-source data and reward model using 4 NFE.

---

🌟 Inference

8 step Z-Image-Turbo generation (distilled by Decoupled-DMD and DMDR)

See Z-Image repo.

🌠 Training

We now only have access to open the training demo code of ImageNet, hoping it can help the community to understand DMDR!

Refer to SiT (class-conditional generation on ImageNet) for training few-step SiT diffusion model.

🥂 Other Research of Our Team

• Decoupled DMD: Rethinking how DMD works and revealing a functional decoupling strategy with CFG Augmentation (CA) as the primary engine for few-step conversion and Distribution Matching (DM) as the regularizer.

🤝🏻 Acknowledgement

This code is mainly built upon DMD2, SRA, ReFL, repositories. Thanks for their contributions to the community.

We also sincerely thank the opensource weights from REPA, DINOv2 and so on. We only use these weights and data for research purpose.

🌺 Citation

If you find DMDR useful, please kindly cite our paper:

@article{jiang2025distribution,
  title={Distribution Matching Distillation Meets Reinforcement Learning},
  author={Jiang, Dengyang and Liu, Dongyang and Wang, Zanyi and Wu, Qilong and Li, Liuzhuozheng and Li, Hengzhuang and Jin, Xin and Liu, David and Li, Zhen and Zhang, Bo and others},
  journal={arXiv preprint arXiv:2511.13649},
  year={2025}
}