An artificial intelligence tool can use brain scans to recreate images people looked at, Live Science reported. The system, called Brain-IT, uses functional magnetic resonance imaging to measure activity in the visual cortex, the part of the brain that processes visual input. A computer model predicts both the broad layout of an image and higher-level information about its content, then passes those predictions to an AI image generator. The research was presented at the ICLR 2026 conference, held April 23-25, and publicized by its authors in September. A paper describing the work was posted on the preprint server arXiv.
Brain-IT builds on an approach called the Universal Brain Encoder, which starts with an image and predicts the fMRI response expected in individual voxels, the small 3-D units in a brain scan. "The key idea is that every voxel, a tiny cube of brain tissue, gets its own learned vector that captures what it cares about in images," said first author Roman Beliy, a doctoral candidate at the Weizmann Institute of Science in Israel. Some voxels may be tuned to edges, textures, or locations; others to higher-level categories such as faces, bodies, or scenes. Traditional image-to-fMRI systems have generally been custom trained for each participant and dataset. The Universal Brain Encoder, the study says, can learn from data from multiple people and transfer across participants, datasets, or MRI systems of different field strengths.
Brain-IT works in the opposite direction: it starts with measured brain activity and reconstructs what the person saw. It sorts thousands of visual-cortex measurements into 128 groups, lets those groups share clues, and produces two kinds of predictions, semantic and structural. The structural prediction gives a generator a starting point; the semantic information guides a diffusion model as it removes remaining static. The authors said Brain-IT surpassed prior leading methods on standard reconstruction metrics, and that one hour of data from a new participant produced results as accurate as earlier approaches trained on 40 hours of recordings. They also warned against treating a realistic output as a perfect copy. In the final stage, an off-the-shelf diffusion generator "beautifies" the image and can introduce details that do not necessarily come from the brain data itself. The scan constrains the picture; the generator supplies some of its realism from patterns learned in photographs.
"We're not reading people's thoughts," co-author Michal Irani, a computer scientist at Weizmann, told Live Science. "The images we do manage to recover from fMRI require people's cooperation and full intended collaboration." The system cannot reconstruct memories or imagined scenes, because those do not offer the same clean visual-cortex signal as a picture actually shown in an experiment. Another limit is practical: reliable results still require cumbersome, expensive fMRI. The team is experimenting with EEG, Irani said, but that work is "in the very preliminary stages." Possible future uses she mentioned include studying visual processing and, eventually, work on conditions such as MS, PTSD, autism, and psychiatric disorders. Those are research aims, not a claim that a clinical test exists today.