Researchers at the Weizmann Institute of Science have developed Brain-IT, an artificial intelligence model that reconstructs images people have viewed using functional magnetic resonance imaging (fMRI) recordings of brain activity.
The research was presented at the International Conference on Learning Representations in 2026. The team comprises Roman Beliy, Amit Zalcher, Jonathan Kogman, Navve Wasserman and Michal Irani.
Brain-IT uses a Brain Interaction Transformer to group brain voxels, the small units measured by fMRI scans, according to functional similarities. This allows the model to identify shared patterns across different participants.
It predicts semantic features, which capture an image’s content, and structural features, which represent its layout. These predictions guide a diffusion model to generate the reconstructed image.
A key finding is that the model can be adapted to new participants using limited individual data. One hour of fMRI recordings produced results comparable to methods trained on approximately 40 hours of data from a participant.
The researchers also reported that Brain-IT outperformed existing approaches in visual comparisons and standard evaluation metrics. Its design aims to improve the reconstruction of both image content and spatial arrangement.
However, the system reconstructs images participants have already viewed rather than providing unrestricted access to their thoughts. Its reliance on fMRI equipment also limits its use outside controlled research settings.
The findings offer a new approach to studying how the brain processes visual information and how AI can translate these neural patterns into images.
