Reproduce “Discrete Meanflow Training Curriculum” on OpenResearch

About this paper

Chia-Hong Hsu, Frank Wood

The authors introduce DMF, a staged method for fine-tuning a pretrained flow model toward one-step image generation. It begins with flow-matching targets, uses discrete intermediate targets that avoid Jacobian-vector products, then applies continuous MeanFlow training. On unconditional CIFAR-10, VE-schedule DMF achieves one-step FID 3.36 versus 3.93 for MeanFlow fine-tuning; the comparison is not a matched ablation, and ImageNet can diverge.

Key insights

• A specific discretization of the MeanFlow objective can maintain a consistency property, allowing for significant computational savings by avoiding costly Jacobian-vector products during training.
• A curriculum learning approach, by adaptively reducing the discretization step size, facilitates efficient training, allows for effective initialization with pretrained flow models, and progressively refines model performance.
• There exists an empirical stability ceiling for excessively fine discretizations, particularly in latent spaces, indicating a trade-off between approximation accuracy and optimization robustness that can lead to divergence.

Summary by alphaXiv.