Brain-IT isn’t a camera. It doesn’t snap photos or record video. But it does something eerily close: reconstruct what a person is seeing — directly from their brain activity.
Developed in Prof. Michal Irani’s lab at the Weizmann Institute of Science, Brain-IT is a new AI model trained to translate functional MRI (fMRI) scans into visual reconstructions. Unlike earlier systems that took dozens of hours to calibrate for a single person, Brain-IT achieves usable image reconstruction after just one hour of brain scanning.
The team trained the model using over 70,000 images shown to eight participants while their brains were scanned. Each scan was broken into roughly 40,000 voxels, the 3D equivalent of pixels. And mapped against visual features like color, object type (faces, food, places), and spatial layout. By comparing responses across subjects, the researchers discovered consistent patterns, not just in known visual processing areas like the parahippocampal place area (PPA), but in 128 functional regions.
Some of those regions were already documented. Others were newly identified: for instance, a functional split within the PPA, where one subregion responded selectively to indoor scenes and another to outdoor ones.
Brain-IT uses two linked components: an encoder and a decoder. The encoder predicts how a given image should activate the brain. The decoder reverses the process. Inferring the image from observed brain activity. Training them together sharpened accuracy in both composition and color, addressing key weaknesses in earlier models Irani described as preserving “semantic meaning” but failing on basic visual fidelity.
This isn’t real-time mind reading. It requires high-resolution fMRI, not consumer-grade wearables. And it only works when the subject is actively viewing static images inside a scanner, not imagining, recalling, or dreaming. Still, the leap in speed and cross-subject generalization is significant.
The implications are clinical and scientific. Not commercial. Researchers say Brain-IT could eventually help people with paralysis communicate via decoded visual intent. It also offers neuroscientists a faster, more scalable tool to map how different brains process visual information.
No hardware launch is planned. No product name has been trademarked. There’s no price, no release date, and no roadmap to a consumer version. What exists is a peer-reviewed method, published via academic channels, validated on eight people, and now documented by outlets including PetaPixel and MIT Technology Review.
That said, the timing matters. As neural interface startups race toward FDA clearance and major tech firms file patents on non-invasive brain-computer interfaces, Brain-IT adds concrete evidence that decoding visual perception is becoming faster, more portable, and more generalizable across individuals.
Next steps? Scaling beyond eight participants. Testing dynamic stimuli. Like short video clips. And validating whether the 128-region framework holds across broader demographics, including people with neurological conditions.
For now, Brain-IT remains a lab-built proof point: not a gadget, not a camera, but a new kind of lens on how vision becomes thought.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.