MixFlow: Mixed Source Distributions Improve Rectified Flows
A training strategy that mixes conditional and unconditional source distributions to improve rectified-flow sampling.
Overview
MixFlow combines conditional and unconditional source distributions when training rectified flows. The approach aims to straighten generation paths and improve the trade-off between sampling speed and image quality.
The repository contains training, inference, and curvature-evaluation tools, with experiment configurations for CIFAR-10, FFHQ, and AFHQv2. Experiments use PyTorch Lightning and Hydra.
Resources
Code · Paper · Workshop submission
Venue
ICLR 2026, 2nd DeLTa Workshop.