Cross-scale Aligned Supervision for Training GANs

Researchers proposed CAT, a transformer-based GAN training method for cross-scale supervision in image generation.

The paper argues that standard intermediate adversarial losses can misalign sample trajectories across scales and cause drift.

CAT preserves scale-wise discrimination while adding generator consistency regularization so intermediate outputs match the final output.

This improves sample consistency, reduces generation error, and supports strong one-step inference on datasets like ImageNet-256.