Distribution Matching Distillation Meets Reinforcement Learning

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Jiang, Dengyang, Liu, Dongyang, Wang, Zanyi, Wu, Qilong, Li, Liuzhuozheng, Li, Hengzhuang, Jin, Xin, Liu, David, Lu, Changsheng, Li, Zhen, Zhang, Bo, Wang, Mengmeng, Hoi, Steven, Gao, Peng, Yang, Harry
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866908911278227456
author Jiang, Dengyang
Liu, Dongyang
Wang, Zanyi
Wu, Qilong
Li, Liuzhuozheng
Li, Hengzhuang
Jin, Xin
Liu, David
Lu, Changsheng
Li, Zhen
Zhang, Bo
Wang, Mengmeng
Hoi, Steven
Gao, Peng
Yang, Harry
author_facet Jiang, Dengyang
Liu, Dongyang
Wang, Zanyi
Wu, Qilong
Li, Liuzhuozheng
Li, Hengzhuang
Jin, Xin
Liu, David
Lu, Changsheng
Li, Zhen
Zhang, Bo
Wang, Mengmeng
Hoi, Steven
Gao, Peng
Yang, Harry
contents Distribution Matching Distillation (DMD) facilitates efficient inference by distilling multi-step diffusion models into few-step variants. Concurrently, Reinforcement Learning (RL) has emerged as a vital tool for aligning generative models with human preferences. While both represent critical post-training stages for large-scale diffusion models, existing studies typically treat them as independent, sequential processes, leaving a systematic framework for their unification largely unexplored. In this work, we demonstrate that jointly optimizing these two objectives yields mutual benefits: RL enables more preference-aware and controllable distillation rather than uniformly compressing the full data distribution, while DMD serves as an effective regularizer to mitigate reward hacking during RL training. Building on these insights, we propose DMDR, a unified framework that incorporates Reward-Tilted Distribution Matching optimization alongside two dynamic distillation training strategies in the initial stage, followed by the joint DMD and RL optimization in the second stage. Extensive experiments demonstrate that DMDR achieves state-of-the-art visual quality and prompt adherence among few-step generation methods, even surpassing the performance of its multi-step teacher model.
format Preprint
id arxiv_https___arxiv_org_abs_2511_13649
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Distribution Matching Distillation Meets Reinforcement Learning
Jiang, Dengyang
Liu, Dongyang
Wang, Zanyi
Wu, Qilong
Li, Liuzhuozheng
Li, Hengzhuang
Jin, Xin
Liu, David
Lu, Changsheng
Li, Zhen
Zhang, Bo
Wang, Mengmeng
Hoi, Steven
Gao, Peng
Yang, Harry
Computer Vision and Pattern Recognition
Distribution Matching Distillation (DMD) facilitates efficient inference by distilling multi-step diffusion models into few-step variants. Concurrently, Reinforcement Learning (RL) has emerged as a vital tool for aligning generative models with human preferences. While both represent critical post-training stages for large-scale diffusion models, existing studies typically treat them as independent, sequential processes, leaving a systematic framework for their unification largely unexplored. In this work, we demonstrate that jointly optimizing these two objectives yields mutual benefits: RL enables more preference-aware and controllable distillation rather than uniformly compressing the full data distribution, while DMD serves as an effective regularizer to mitigate reward hacking during RL training. Building on these insights, we propose DMDR, a unified framework that incorporates Reward-Tilted Distribution Matching optimization alongside two dynamic distillation training strategies in the initial stage, followed by the joint DMD and RL optimization in the second stage. Extensive experiments demonstrate that DMDR achieves state-of-the-art visual quality and prompt adherence among few-step generation methods, even surpassing the performance of its multi-step teacher model.
title Distribution Matching Distillation Meets Reinforcement Learning
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2511.13649