Daily Paper Briefing

Abstract Overview

This paper studies Representation Distribution Matching (RDM) as a framework for training one-step visual generators by matching generated and reference feature distributions under frozen pretrained encoders. It organizes the method space along two design axes: how distributions are compared and which representations are used for comparison. Through controlled studies, the authors argue that MMD becomes effective when estimated with an exact within-batch repulsion term and a Nyström attraction to a frozen full-data reference, that large fresh generated batches are important, and that matching only a single encoder can be gamed. These choices are combined into improved RDM (iRDM), which is evaluated on one-step ImageNet generation and on post-training FLUX.2 [klein] from four steps to one step for text-to-image generation.

Novelty

The paper’s main novelty is a systematic design-space analysis of representation distribution matching for one-step generation, rather than presenting only a single heuristic recipe. It also introduces iRDM, which combines Nyström-based MMD estimation, large fresh batches, joint image-text matching for conditional tasks, and constrained balancing across multiple encoders, alongside a multi-encoder evaluation metric based on sliced Wasserstein distance.

Results

On ImageNet-256, iRDM achieves a reported one-step state of the art with SW_r^14 = 1.30 and is preferred over the prior best one-step generator on 71.2% of matched samples by PickScore. In text-to-image post-training, the same recipe converts four-step FLUX.2 [klein] into a one-step model that surpasses the four-step version on GenEval (0.826 vs. 0.794) and PickScore (22.76 vs. 22.58) in about 90 H200 GPU-hours. Ablations further support the roles of Nyström-MMD, large fresh batches, joint image-text matching, and constrained multi-encoder weighting.

Key Points

1. The paper identifies two key RDM design axes: distribution comparison strategy and representation choice, and studies them with controlled ablations.
2. It argues that single-encoder matching is insufficient because generators can overfit one representation while remaining visibly flawed, motivating balanced optimization across multiple encoders.
3. The resulting iRDM recipe improves one-step image generation quality on ImageNet and enables one-step post-training of FLUX.2 [klein] with competitive or better prompt-following and preference metrics.

References

• arXiv: https://arxiv.org/abs/2607.02375v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2607.02375v1
• Hugging Face Papers: https://huggingface.co/papers/2607.02375
• Project: https://alan-lanfeng.github.io/rdm/