This research paper examines artifacts in vision transformer feature maps, specifically high-norm tokens appearing in non-informative image areas. The authors propose adding "register" tokens to the input sequence as a solution. This simple addition eliminates the artifacts, improves performance on dense prediction tasks and object discovery, and results in smoother feature and attention maps. The findings apply to both supervised and self-supervised vision transformer models, significantly enhancing their interpretability and effectiveness. Experiments across various models and tasks validate the approach's efficacy and generalizability.

ai , artificial intelligence , arxiv , research , paper , publication , llm, genai, generative ai , large visual models, large language models, large multi modal models, nlp, text, machine learning, ml, nividia, openai, anthropic, microsoft, google, technology, cutting-edge, meta, llama, chatgpt, gpt, elon musk, sam altman, deployment, engineering, scholar, science, apple, samsung, anthropic, turing