# Timothée Darcet Lead author of "Vision Transformers Need Registers". Topic: [[Topics/LLMs/LLMs|LLMs]] · [[Topics/LLMs/Sources|Sources]] - Lead author, with Maxime Oquab, Julien Mairal and Piotr Bojanowski (Meta FAIR and Inria), of "Vision Transformers Need Registers" (arXiv Sep 2023; ICLR 2024) ([arXiv](https://arxiv.org/abs/2309.16588)). - Finding: large ViTs (DINOv2, OpenCLIP, DeiT-III) take over low-information background patches as high-norm "artifact" tokens that hold global information. Adding a few extra learnable **register** tokens gives the model a place to put that information and removes the artifacts. - The post mixes up the two: Meta found the artifacts; registers were the fix.