Distance Marching for Generative Modeling (arxiv.org)

A new approach called Distance Marching has been introduced to enhance time-unconditional generative models, which traditionally struggle with disambiguation of noise levels and denoising directions in the absence of temporal input. This technique employs innovative inference methods alongside novel loss functions designed to prioritize closer targets, resulting in more effective denoising pathways that are better aligned with the data manifold. Significantly, Distance Marching has been shown to improve the Fréchet Inception Distance (FID) scores by 13.5% on benchmark datasets such as CIFAR-10 and ImageNet compared to existing time-unconditional models. It also outperforms flow matching techniques in class-conditional generation of ImageNet, achieving superior FID with significantly fewer sampling steps—60% less on average—across various architecture sizes. Additionally, the method's distance predictions facilitate early stopping during sampling and assist in out-of-distribution (OOD) detection, underscoring its potential as a foundational element in future generative modeling strategies.