High-fidelity single-step generation under limited data

Establish high-fidelity single-step generation for MeanFlow-based generative models under limited target-domain data, addressing the failure of current MeanFlow-Transfer and Continuous Adversarial MeanFlow methods to recover the one- to two-step quality achieved by MeanFlow models in data-rich source domains.

Background

The paper introduces MeanFlow-Transfer (MF-T) to adapt heterogeneous pretrained diffusion and flow generators into few-step MeanFlow generators on new domains with limited data, and Continuous Adversarial MeanFlow (CAMF) to improve perceptual quality through adversarial post-training. Although the combined method substantially accelerates source models and achieves strong results at four sampling steps, the authors report that it does not recover the one- to two-step quality attained by MeanFlow models trained in data-rich source domains. They therefore identify high-fidelity single-step generation under limited data as an unresolved challenge.