Cross-Attention Makes Inference Cumbersome in Text-to-Image Diffusion Models

arXiv:2404.02747v1 Announce Type: new
Abstract: This study explores the role of cross-attention during inference in text-conditional diffusion models. We find that cross-attention outputs converge to a fixed point after few inference steps. Accordingly, the time point of convergence naturally divides the entire inference process into two stages: an initial semantics-planning stage, during which, the model relies on cross-attention to plan text-oriented visual semantics, and a subsequent fidelity-improving stage, during which the model tries to generate images from previously planned semantics. …

arxiv attention cs.cv diffusion diffusion models image image diffusion inference text text-to-image type