ReVersion: Diffusion-Based Relation Inversion from Images
Ziqi Huang, Tianxing Wu, Yuming Jiang, Kelvin C. K. Chan, Ziwei Liu
Abstract
ReVersion "Michael Jackson <R> wall" "cat <R> canvas" "Spiderman <R> wall" "hamburger <R> vase" "cat <R> building" exemplar images <R> ReVersion "cat <R> cat" "panda <R> panda" "otter <R> otter" "rabbit <R> child" "Batman <R> Batman" exemplar images <R> ReVersion "Spiderman <R> basket" "hamster <R> paper bag" "sea <R> cup" "panda <R> pot" "rabbit <R> cup"Fig. 1.We propose a new task, Relation Inversion: Given a few exemplar images, where a relation co-exists in every image, we aim to find a relation prompt R to capture this interaction, and apply the relation to new entities to synthesize new scenes.The above images are generated by our ReVersion Framework.
Topics & Concepts
ReversionInversion (geology)Mean reversionDiffusionRelation (database)Computer scienceGeologyMathematicsData miningEconometricsSeismologyPhysicsThermodynamicsTectonicsBiochemistryPhenotypeGeneChemistryMultimodal Machine Learning ApplicationsDomain Adaptation and Few-Shot LearningGenerative Adversarial Networks and Image Synthesis