Mind-the-Glitch: Visual Correspondence for Detecting Inconsistencies in Subject-Driven Generation
Mind-the-Glitch introduces a method to find visual and semantic connections between parts of images generated by diffusion models. It separates model features into distinct semantic and visual components, uses this separation to train the model better, and creates a new metric called Visual Semantic Matching (VSM) to precisely measure inconsistencies in subject-driven image synthesis.