表征分布匹配用于单步视觉生成

摘要

我们阐明了表示分布匹配（RDM，即通过冻结预训练编码器匹配生成特征分布与参考特征分布来训练一步图像生成器的范式）的设计空间。识别出两个设计轴——分布的比较方式以及用于比较的表示空间，并沿这两个轴开展受控研究，得出三项发现。首先，经典的最大均值差异（MMD）在十年前无法训练出令人信服的生成器，但如今若正确估计，则成为一个强大且可扩展的目标函数。其次，生成批次成为关键操作变量，其最优值超过2048，远超常规批次大小。第三，任何单一表示都可能被操纵，即图像在视觉上仍显虚假而该表示的得分却低于真实得分，因此我们针对平衡的编码器组合进行匹配，并使用SW_r14（一种基于14个编码器的切片瓦瑟斯坦距离，独立于训练损失且抗操纵性）进行评估。结合优选方案得到改进型RDM（iRDM）：它在ImageNet上以SW_r14 1.30的指标确立了单步生成器的当前最优性能，并通过PickScore（一种我们目标函数从未优化过的人类偏好代理指标）得到验证——在71.2%的匹配样本中，它优于先前最优的单步生成器。相同的方案将四步FLUX.2 [klein]后训练为单步生成器，在GenEval（0.826 vs 0.794）和PickScore（22.76 vs 22.58）上均超越四步版本，耗时90 H200 GPU小时。项目页面：https://alan-lanfeng.github.io/rdm/。

English

We elucidate the design space of Representation Distribution Matching (RDM), our name for the paradigm that trains a one-step image generator by matching generated and reference feature distributions under frozen pretrained encoders. We identify two design axes, how the distributions are compared and the representations they are compared in, and controlled studies along them yield three findings. First, the classical MMD, which could not train convincing generators a decade ago, becomes a strong and scalable objective once estimated right. Second, the generated batch is then the operative variable, with an optimum above 2048, far beyond customary batch sizes. Third, any single representation can be gamed, driven below the real score while images stay visibly fake, so we match against a balanced battery of encoders and evaluate with SW_r14, a Sliced-Wasserstein distance over 14 encoders that is independent of the training loss and resists gaming. Combining the preferred choices yields improved RDM (iRDM): it sets the one-step state of the art on ImageNet at SW_r14 1.30, corroborated by PickScore, a human-preference proxy our objective never optimizes, which prefers it over the prior best one-step generator on 71.2% of matched samples. The same recipe post-trains the four-step FLUX.2 [klein] into a one-step generator, surpassing the four-step version on GenEval, 0.826 to 0.794, and on PickScore, 22.76 to 22.58, in 90 H200 GPU-hours. Project page: https://alan-lanfeng.github.io/rdm/.