Discrete Meanflow Training Curriculum
2026-04-10 • Machine Learning
Machine Learning
AI summaryⓘ
The authors study image generation models that can make pictures in just one or a few steps. Usually, models that take many steps are stable and make good images, but one-step models are harder to train and less stable. They focus on Meanflow models, which are good at few-step and one-step image generation but usually need a lot of time and data to train. By introducing a new training method called Discrete Meanflow (DMF) Curriculum, they show how to train these models much faster and with less data, starting from an existing Flow Model. Their method achieves strong results quickly on the CIFAR-10 dataset.
Flow-based generative modelsMeanflow modelsOne-step samplingMulti-step samplingTraining curriculumDiscrete Meanflow (DMF)FID scoreCIFAR-10 dataset
Authors
Chia-Hong Hsu, Frank Wood
Abstract
Flow-based image generative models exhibit stable training and produce high quality samples when using multi-step sampling procedures. One-step generative models can produce high quality image samples but can be difficult to optimize as they often exhibit unstable training dynamics. Meanflow models exhibit excellent few-step sampling performance and tantalizing one-step sampling performance. Notably, MeanFlow models that achieve this have required extremely large training budgets. We significantly decrease the amount of computation and data budget it takes to train Meanflow models by noting and exploiting a particular discretization of the Meanflow objective that yields a consistency property which we formulate into a ``Discrete Meanflow'' (DMF) Training Curriculum. Initialized with a pretrained Flow Model, DMF curriculum reaches one-step FID 3.36 on CIFAR-10 in only 2000 epochs. We anticipate that faster training curriculums of Meanflow models, specifically those fine-tuned from existing Flow Models, drives efficient training methods of future one-step examples.