Israeli AI System Recreates Visual Images Directly from Brain Scans

Brain-IT, an artificial-intelligence framework created in Professor Michal Irani’s laboratory at the Weizmann Institute of Science, converts fMRI data into images that closely resemble what a subject saw. The model learns from a large dataset of brain recordings collected from eight volunteers, grouping similar neural responses into 128 functional clusters that generalize across individuals. Using an encoder to predict brain activity and a decoder trained on synthetic responses, the system produces a coarse layout and detailed content cues, which an AI image generator refines into a final picture.

In tests on 1,000 unseen photographs, Brain-IT achieved the highest average scores on seven of eight standard image-accuracy metrics, beating competing models even with only an hour of personal data, and sometimes as little as fifteen minutes. The researchers suggest the technology could eventually aid studies of visual perception, consciousness disorders, and communication for non-verbal patients, though clinical applications remain distant.