Publication
Better Source, Better Flow: Learning Condition-Dependent Source Distribution for Flow Matching
A method that learns condition-dependent source distributions for flow matching, enabling straighter transport paths and up to 3x faster convergence.
Abstract
We introduce a method for learning condition-dependent source distributions in flow matching frameworks. Standard flow matching uses a fixed Gaussian as the source distribution, leading to curved transport paths and slow convergence. By adapting the source distribution based on the target condition, our approach achieves straighter trajectories, reduced variance, and up to 3x faster convergence in FID compared to standard Gaussian source distributions.