Researchers at the Weizmann Institute of Science built Brain-IT, an artificial intelligence system that reconstructs images from brain scans, opening avenues for paralyzed patients while renewing debates on mental privacy.
A team of scientists led by Michal Irani at the Weizmann Institute of Science in Israel built an AI decoder that uses functional magnetic resonance imaging data to recreate external images viewed by human subjects. The system, detailed in a paper submitted to the International Conference on Learning Representations, earned acceptance at ICLR 2026. Alongside Irani, the researchers include Roman Beliy, Amit Zalcher, Jonathan Kogman, and Navve Wasserman.
Earlier brain-to-image systems could capture the general semantic subject of a picture but frequently faltered on details like color and composition. There exist nowadays models that translate brain activity into images, and they can even produce impressive reconstructions that preserve the semantic meaning of the image reasonably well,
Irani said in a statement provided by the institute. However, they tend to make mistakes in basic features such as composition and color. The new model we developed outperforms them in reconstructing both the content of the image and its details.
Brain-IT Architecture Processes 40,000 Voxels
The system centers on an architecture called a Brain-Interaction Transformer, or BIT. This network models the brain as 40,000 voxels—three-dimensional pixels—and records how each voxel responds as a subject observes a picture, logging features like color, location, and higher-level categories such as whether a person is looking at a face or a sandwich. The architecture relies on two branches: one predicts an image’s structure and color placement, while the other identifies its semantic content. Those predictions then feed into a diffusion model to generate the final reconstruction.
Training an AI on brain scans typically demands vast amounts of data, which remain scarce because recording fMRI sessions is both tedious and expensive.
To bypass this bottleneck, the Weizmann researchers trained their model using existing data from eight volunteers who viewed roughly 9,000 images each while undergoing high-resolution fMRI scans, drawing from the Natural Scenes Dataset.
To expand their training pool without putting participants through endless hours in a scanner, the team built a universal brain encoder capable of predicting what brain activity looks like given an image, effectively reversing the process. This allowed the system to train on pictures never actually viewed by anyone inside an fMRI machine. We realized that by translating back and forth – from a random image that had never been viewed in an fMRI machine, to a predicted brain scan, and then back to the image we started with – the models would effectively build themselves a massive dataset,
Irani explained in a statement.
Brain-IT Cuts Calibration Time to One Hour
Brain-IT adapts to a new person in about one hour, cutting calibration time down from approximately 40 hours of fMRI recordings demanded by older brain-decoding methods.
“None of us can afford 40 hours of imaging for a new subject,” UC Santa Barbara neuroscientist Tommy Sprague told the Review. “It’s something like $600 to $1,000 an hour.”
Tommy Sprague, UC Santa Barbara neuroscientist
Despite the speed improvement, the technology remains bound by the physical limitations of fMRI equipment, which is bulky, expensive, and requires a controlled laboratory setting.
The system does not always succeed. Irani noted that the model makes occasional errors, such as interpreting a picture of a cake as three sandwiches or turning a dog sitting in a bathtub into a similarly colored goat in the same setting.
The system also cannot decode internal thoughts, memories, or intentions. If we overcome all these obstacles, it’s possible that in the future we may even be able to read dreams,
Irani stated.
Brain-IT Maps 128 Shared Brain Regions
While training on the dataset, Brain-IT mapped 128 functional brain regions shared across participants.
The primary therapeutic goal for the technology is assisting completely paralyzed or “locked-in” patients who cannot speak or move.
External Researchers Warn of Long-Term Risks
External researchers have welcomed the technical achievement while warning of long-term risks if brain-decoding technology migrates from bulky fMRI machines to more portable and accessible hardware, such as EEG scalp sensors.
The results seem very impressive,
Sprague said. But if there’s a way to surreptitiously extract information about what you’re thinking about, then …150 years of sci-fi can come true anytime, and that’s worrisome in a lot of ways.
While investigators are currently exploring EEG-based decoding, current studies have not established that portable scalp sensors can match the precision demonstrated by fMRI systems.
