RDM

Representation Distribution Matching (RDM, with the trained generator referred to as iRDM) is a research method for training one-step image generators. It matches the distributions of generated and real images in the feature spaces of frozen pretrained encoders. It uses no online teacher, no adversary and no sampling trajectory, so a single network evaluation produces an image.

The method varies two choices: how distributions are compared and in which spaces. For comparison, it uses MMD with an exact within-batch repulsion and a Nyström attraction to a frozen full-data reference, with very large generated batches (above 2048) enabled by gradient caching. For conditional tasks it matches the joint image-text distribution. For the spaces, it uses a battery of frozen encoders balanced by a proportional Lagrangian controller, to avoid gaming any single encoder.

The authors report state-of-the-art one-step ImageNet generation. They also post-train the four-step FLUX.2 [klein] into a one-step model in about 90 H200 GPU-hours. They report GenEval 0.826 and PickScore 22.76, against 0.804 and 22.36 for a one-step DMD2 distillation of the same teacher. A demo of the one-step FLUX.2 [klein] model is said to run on Hugging Face Spaces. A video creator noted that quality beats other one-step models but still lacks some detail. The paper is by Feng, Li, Zablocki, Cord and Alahi (arXiv 2607.02375).