Global transport improves guided image generation across models and conditions
Global Transport Couplings for Classifier-Free Guided Flows
Machine Learning
Summary
Generating images from descriptions or categories can be tricky when the computer has many conditions to consider. The authors study a way to link the initial noise used in image generation in a way that doesn’t depend on specific categories, called Global Transport (GT). On its own, GT can make image generation worse, but when combined with a technique called classifier-free guidance, it helps generate better images consistently. This means it's important to test new methods both with and without guidance to understand their true effects.
What this means in practice
- •For image generation teams: Improve the quality of text- or class-conditioned image generation by integrating global transport couplings with classifier-free guidance.
- •For machine learning engineers: Enhance unconditional flow models’ training stability by applying global optimal-transport couplings independent of class labels.
Authors
Katarina Petrović, Zander W. Blasingame, Danyal Rehman, İsmail İlkan Ceylan, Michael Bronstein, Stephen Y. Zhang, Lazar Atanackovic, Alexander Tong
Abstract
Optimal-transport couplings have been shown to reduce training variance in unconditional flow models, but their role in conditional generation remains unclear. A natural approach constructs separate couplings for each condition, but this is impractical for large or continuous conditioning spaces found in modern image foundation models. We introduce Global Transport (GT), a global class-agnostic optimal-transport coupling, computed without class labels. GT can associate different conditions with different regions of the source noise, and consequently worsens performance without guidance. However, when combined with classifier-free guidance (CFG), GT consistently improves generation across domains, model scales, and sampling budgets. This reversal suggests that couplings for conditional flows should be evaluated both empirically and theoretically under the guided flow used at inference, rather than on unguided generation. We evaluate GT over both discrete class and continuous text conditioned image generation across model scales, and investigate how coupling choice alters guided trajectories. These results identify coupling design in the guided flow setting as a simple training time axis to improve performance without modifying existing architectures, samplers, or guidance mechanisms.