Distance Marching for Generative Modeling
7 days ago
- #Distance Marching
- #Machine Learning
- #Generative Modeling
- Proposes Distance Marching, a new time-unconditional approach for generative modeling.
- Introduces two principled inference methods and losses focusing on closer targets for better denoising directions.
- Improves FID by 13.5% on CIFAR-10 and ImageNet over recent time-unconditional baselines.
- Surpasses flow matching in class-conditional ImageNet generation despite removing time input.
- Achieves lower FID with fewer sampling steps and better performance across backbone sizes.
- Distance prediction aids in early stopping during sampling and OOD detection.
- Suggests distance field modeling as a principled lens for generative modeling.