Tag
This paper investigates how semantic information is distributed across textual tokens in text-to-image models, finding that information concentration and cross-item interactions significantly affect image generation alignment. The authors use patching techniques to demonstrate that simple encoding-stage interventions can improve alignment quality.