Distance Marching for Generative Modeling
Distance Marching is a new time-unconditional approach for generative modeling inspired by distance field modeling, designed to improve denoising direction accuracy. It introduces principled inference methods and losses focused on closer targets, leading to better alignment with the data manifold. The method achieves superior performance on CIFAR-10 and ImageNet, with faster sampling and improved FID scores compared to flow matching.
- ▪Distance Marching improves FID by 13.5% on CIFAR-10 and ImageNet over recent time-unconditional baselines.
- ▪For class-conditional ImageNet generation, it surpasses flow matching despite removing time input.
- ▪The approach achieves lower FID using only 60% of the sampling steps and shows benefits in early stopping and out-of-distribution detection.
- ▪Distance prediction in this method aids both sampling efficiency and anomaly detection.
- ▪The paper was submitted on February 3, 2026, and authored by Zimo Wang and six others.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2602.02928 (cs) [Submitted on 3 Feb 2026] Title:Distance Marching for Generative Modeling Authors:Zimo Wang, Ishit Mehta, Haolin Lu, Chung-En Sun, Ge Yan, Tsui-Wei Weng, Tzu-Mao Li View a PDF of the paper titled Distance Marching for Generative Modeling, by Zimo Wang and 6 other authors View PDF HTML (experimental) Abstract:Time-unconditional generative models learn time-independent denoising vector fields. But without time conditioning, the same noisy input may correspond to multiple noise levels and different denoising directions, which interferes with the supervision signal. Inspired by distance field modeling, we propose Distance Marching, a new time-unconditional approach with two principled inference methods.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv.org.
Discussion
0 commentsMore from arXiv.org
