Amortized Moment Matching for Visual Generation

Researchers introduced amortized moment matching, a neural method that learns distributional moments directly with networks instead of explicit calculations.

The approach uses an Amortized Fréchet Distance loss and scales better to high-dimensional visual generation settings.

Reported results show improved training stability, stronger one-step image generation, and better text-to-image performance on benchmarks such as ImageNet and GenEval.

The method also showed gains in instruction following and scoring metrics including FLUX.2 and PickScore.