Rethinking Global Text Conditioning in Diffusion Transformers

A new paper accepted at ICLR 2026 reveals that the pooled text embedding in diffusion transformers like Stable Diffusion, often considered redundant, can be repurposed as a powerful guidance signal. This training-free method incurs negligible runtime overhead but enables controllable shifts toward more desirable outputs. It brings significant improvements across text-to-image, text-to-video generation, and image editing tasks, proving a core architectural assumption wrong and opening new optimization paths.