This paper shows a simple way to make image-generating AIs (diffusion Transformers) produce clearer, more accurate pictures by letting the model guide itself from the inside.
Normalizing Flows are models that learn how to turn real images into simple noise and then back again.