Label-free optimal transport reportedly outperforms class-conditional matching at higher guidance scales
A post announcing the preprint says the method reached 1.91 FID on ImageNet-256 at higher classifier-free guidance scales.
TLDR
A preprint announcement says Global Transport (GT), a label-free optimal transport coupling, beats class-conditional and independent couplings at higher classifier-free guidance (CFG) scales, reaching 1.91 FID on ImageNet-256. A separate user says matching training examples to the closest noise samples in each batch can improve unconditional flow models but is impractical for conditional ones; ignoring conditioning for the matching still improves performance with CFG.
Combined views
720
1 Source, first seen ago
