TexTailor: Inference-Time Textual Guidance Tailoring for Multimodal Diffusion Transformers
TexTailor is a novel inference-time method for Multimodal Diffusion Transformers that adjusts textual guidance block-by-block without further training. By systematically analyzing how different model blocks process text, it improves text alignment for image generation, enables precise editing, and accelerates inference. This technique boosts performance on benchmarks like SD3.5 by selectively enhancing textual information in crucial parts of the model.