The method combines exact within-batch repulsion with a Nyström attraction to a frozen reference distribution and uses large fresh generation batches with gradient caching. It balances fourteen encoder representations, including image and text features for conditional generation, so the model cannot optimize a single representation while leaving visible artifacts in other spaces.
RDM is useful for researchers and infrastructure teams working on fast text-to-image generation. The release shows one-step ImageNet generation with an SW_r14 distance of 1.30 and post-training of four-step FLUX.2 [klein] into a one-step model that exceeds the teacher on GenEval and PickScore, with public code and checkpoints for reproduction.
