Lost in Translation: Latent Concept Misalignment in Text-to-Image Diffusion Models
Published in ECCV, 2024
This paper discovers and tackles a special type of misalignment issue in text-to-image diffusion models named as Latent Concept Misalignment (LC-Mis). Main contributions includes demonstration of the issue, collection of the LC-Mis dataset and the method Mixture of Concept Experts (MoCE).