Presentation

· Contributors · Organizations · Search Program

Cross-image Attention for Zero-shot Appearance Transfer

DescriptionOur zero-shot appearance transfer approach leverages the semantic understanding of text-to-image models to transfer visual appearance across semantically similar objects. Using our cross-image attention, we establish semantic correspondences between two images, combining one image's structure with another's appearance during generation. This allows transferring appearance across diverse inputs with no model training.

Authors