arXiv 2507.185692025-07-24cs.CV

面向高效图像与视频合成的扩散蒸馏对抗分布匹配

Adversarial Distribution Matching for Diffusion Distillation Towards Efficient Image and Video Synthesis

Sun Yat-Sen University（中山大学）；ByteDance Seed Vision（字节跳动种子视觉）；Guangdong Provincial Key Laboratory of Information Security Technology（广东省信息安全技术重点实验室）；Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China（教育部机器智能与高级计算重点实验室）；Pazhou Lab (HuangPu), Guangzhou, China（琶洲实验室（黄埔），广州，中国）

Yanzuo Lu, Yuxi Ren, Xin Xia, Shanchuan Lin, Xing Wang, Xuefeng Xiao, Andy J. Ma, Xiaohua Xie, Jian-Huang Lai

备注：: Accepted by ICCV 2025 (Highlight)

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对 DMD 因逆 KL 散度导致的模式崩溃问题，本文提出 ADM 对抗分布匹配及 DMDX 流程，在图像与视频生成模型上实现更高效的单步和多步蒸馏。

摘要：

Distribution Matching Distillation（DMD，分布匹配蒸馏）是一种很有前景的分数蒸馏技术，可将预训练的教师扩散模型压缩为高效的单步或多步学生生成器。然而，它依赖于逆 Kullback-Leibler（KL）散度最小化，在某些应用中可能导致模式崩溃（或模式寻找）。为规避这一固有缺陷，我们提出 Adversarial Distribution Matching（ADM，对抗分布匹配），这是一个新颖框架，利用基于扩散的判别器，以对抗方式对齐真实与伪分数估计器之间的潜变量预测，用于分数蒸馏。在极具挑战性的单步蒸馏场景中，我们进一步通过在潜空间和像素空间中使用混合判别器的对抗蒸馏来改进预训练生成器。与 DMD2 预训练中使用的均方误差不同，我们的方法在从教师模型收集的 ODE 对上引入分布损失，从而为下一阶段的分数蒸馏微调提供更好的初始化。通过将对抗蒸馏预训练与 ADM 微调结合到名为 DMDX 的统一流程中，所提方法在 SDXL 上取得了优于 DMD2 的单步性能，同时消耗更少的 GPU 时间。此外，在 SD3-Medium、SD3.5-Large 和 CogVideoX 上应用多步 ADM 蒸馏的实验，为高效图像和视频合成设立了新的基准。

英文摘要：

Distribution Matching Distillation (DMD) is a promising score distillation technique that compresses pre-trained teacher diffusion models into efficient one-step or multi-step student generators. Nevertheless, its reliance on the reverse Kullback-Leibler (KL) divergence minimization potentially induces mode collapse (or mode-seeking) in certain applications. To circumvent this inherent drawback, we propose Adversarial Distribution Matching (ADM), a novel framework that leverages diffusion-based discriminators to align the latent predictions between real and fake score estimators for score distillation in an adversarial manner. In the context of extremely challenging one-step distillation, we further improve the pre-trained generator by adversarial distillation with hybrid discriminators in both latent and pixel spaces. Different from the mean squared error used in DMD2 pre-training, our method incorporates the distributional loss on ODE pairs collected from the teacher model, and thus providing a better initialization for score distillation fine-tuning in the next stage. By combining the adversarial distillation pre-training with ADM fine-tuning into a unified pipeline termed DMDX, our proposed method achieves superior one-step performance on SDXL compared to DMD2 while consuming less GPU time. Additional experiments that apply multi-step ADM distillation on SD3-Medium, SD3.5-Large, and CogVideoX set a new benchmark towards efficient image and video synthesis.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。