arXiv：cs.LG（机器学习，全量分类）· Chi Zhang, Shi Haoyang, Yueyi Liu, Ruichuan An, Junkang Zhou, Chang Li, Xiuyuan Lu, Yichi Zhang, Bo Wang, Yuhang Wu, Sen Cui, Miao Liu·· 15 小时前AI 评分45
统一分布训练框架 MGFlow：一步式视觉生成的新 SOTA
Unifying Distributional Training for One-Step Visual Generation
AI 导读
研究者提出统一理论框架 MGFlow，将分布建模与匹配差异分离，用高斯混合在全局矩与样本表示之间以可调粒度建模特征分布，同时支持最优传输与基于分数的匹配。在 ImageNet 256×256 上，MGFlow 大幅超越 FD-Loss 基线，pMF-H 上达到 1.45 FDr⁶、JiT-H 上 1.64 的 SOTA 结果。
正文
Authors:Chi Zhang, Shi Haoyang, Yueyi Liu, Ruichuan An, Junkang Zhou, Chang Li, Xiuyuan Lu, Yichi Zhang, Bo Wang, Yuhang Wu, Sen Cui, Miao Liu
Abstract:Distributional training provides collective supervision for one-step visual generation by matching real and generated features in frozen representation spaces. We introduce a unified theoretical framework that separates distribution modeling from matching discrepancy and connects global objectives to pointwise feature updates through Wasserstein gradient flow. Under this framework, FD-Loss and Gaussian-kernel Drifting are recovered through Gaussian optimal transport and kernel-density-based KL matching, respectively. The framework motivates MGFlow, which models feature distributions with Gaussian mixtures at an adjustable granularity between global moments and sample-based representations. MGFlow supports both optimal transport and score-based matching, and couples mass-constrained sample assignment with paired component updates to address mode collapse that mixture expressivity alone does not resolve. On ImageNet $256\times256$, MGFlow substantially surpasses the FD-Loss baseline, achieving state-of-the-art results with 1.45 $\mathrm{FDr}^6$ on pMF-H and 1.64 on JiT-H. For text-to-image generation, MGFlow post-trains FLUX.2 [klein] 4B into a one-step generator that outperforms the original four-step model on both GenEval and PickScore.
| Comments: | Project page: this https URL |
| Subjects: | Machine Learning (cs.LG) |
| Cite as: | arXiv:2609.35763 [cs.LG] |
| (or arXiv:2609.35763v3 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2609.35763 arXiv-issued DOI via DataCite |
Submission history
From: Haoyang Shi [view email]
[v1]
Mon, 28 Sep 2026 17:59:14 UTC (34,829 KB)
[v2]
Wed, 30 Sep 2026 17:47:11 UTC (35,680 KB)
[v3]
Thu, 1 Oct 2026 17:58:34 UTC (40,216 KB)
来源：arXiv：cs.LG（机器学习，全量分类） · arxiv.org