We propose to treat LLM activations as images and detect hallucinations more accurately (and efficiently) across LLMs with a vision-inspired architecture.