Plate Nº 39 · recorded November 13, 2026
Neuroscience & MindReported finding
Brain-IT AI Reconstructs Seen Images from fMRI Brain Scans
A new AI system, Brain-IT, turns fMRI scans of the visual cortex into images of what people saw, matching older methods with 40 times less brain-scan data.
By Priya Raman4 min read855 words
In brief
- Brain-IT reconstructed images from just 1 hour of fMRI data as accurately as prior methods trained on 40 hours.
- The system sorts visual-cortex measurements into 128 groups of shared visual information.
- The study was presented at the ICLR 2026 conference, held April 23–25, and posted to arXiv on October 29, 2025.
- The system cannot reconstruct memories or imagined images, only scenes a person actually viewed with full cooperation.
- Researchers are testing EEG-based data collection, but that work is in very preliminary stages.
One hour of brain-scan data from a new participant was enough for a new AI system to reconstruct images as accurately as earlier methods that required 40 hours of recordings. The system, called Brain-IT, converts functional magnetic resonance imaging (fMRI) measurements of the visual cortex (the part of the brain that processes visual input) into pictures of what a person was looking at — and it can also run in reverse, predicting brain activity from an image.
Roman Beliy, a doctoral candidate in the Department of Computer Science and Applied Mathematics at the Weizmann Institute of Science in Israel, presented the work at the ICLR 2026 conference, held April 23–25. The study was uploaded to the preprint database arXiv and publicized by its authors in September.
How does Brain-IT read brain activity?
The system builds on an approach called the Universal Brain Encoder. That model starts with an image and predicts the fMRI response expected in individual "voxels" — small 3D units of brain tissue.
"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," Beliy said.
Some voxels tune into simple elements such as edges, textures or locations. Others respond to abstract categories like faces, bodies or scenes. A voxel uses its learned profile to decide which parts of an image matter to it — essentially asking, "I recognize edges. Which parts of this image are edges?"
Traditional image-to-fMRI systems had to be custom trained for each participant and dataset. The Universal Brain Encoder is more flexible: it learns from multiple people and can transfer information across participants, datasets and MRI systems with different magnetic-field strengths.
Brain-IT inverts this process. Starting from measured brain activity, it asks: "based on this brain activity, what was the subject looking at?" The decoder sorts thousands of tiny brain-scan measurements from the visual cortex into 128 groups, based on the kinds of visual information each group tends to carry. The groups then share clues with one another — combining, for example, a scene's shape, color and possible objects — to build a fuller picture of the image.
What exactly does the model produce?
The decoder outputs two types of predictions:
- Semantic information — the features associated with what is likely in a scene. For a dog standing in grass, this includes the features of a dog and an outdoor, grassy setting.
- Structural information — the broad visual layout. In the same example, the outline of the dog and its position in the frame.
The structural prediction gives the image generator a starting point rather than pure noise. The semantic information then guides a diffusion generator (an AI trained on massive numbers of photographs) that strips away the remaining static until an image emerges.
How accurate are the reconstructions?
Researchers compared Brain-IT's synthetic images with the images participants actually saw. Brain-IT surpassed prior leading methods on standard objective reconstruction metrics and produced more faithful results, the authors wrote.
But a high-scoring or realistic-looking output is not a perfect copy of what someone saw. The final stage relies on an off-the-shelf diffusion generator, which can smooth textures, sharpen facial features or fill in missing details. Beliy said this "beautification" can introduce details that do not necessarily come from the brain data itself. The scan constrains the generated picture; the AI model supplies some of its realism from learned patterns in photographs.
Why is this not mind reading?
The system cannot reproduce memories or imagined scenes.
"We're not reading people's thoughts," study co-author Michal Irani, a computer scientist at the Weizmann Institute of Science, told Live Science. "The images we do manage to recover from fMRI require people's cooperation and full intended collaboration."
"Brain-IT relies on how the brain processes a seen image," Irani said. Information flows from the eyes to the primary visual cortex and onward through visual areas that recognize objects and scenes. In controlled experiments, researchers also know precisely which image was shown, giving the model a clean target during training and evaluation. Memories and imagined images offer no such reference: they can be vague, shaped by memory and thought networks, and reactivate visual regions only partially and weakly.
What could it be used for?
A practical barrier to wide deployment is that reliable results currently require fMRI scanning, which is cumbersome and expensive. The team is experimenting with easier-to-collect electroencephalogram (EEG) data, recorded via electrodes on the head, but Irani said that work is in very preliminary stages.
The system could help brain researchers understand which visual information different brain regions represent, how those regions interact and how individuals' patterns differ. Such tools could eventually reveal differences in visual processing that help clinicians spot neurological, neurodevelopmental and psychiatric conditions.
"There are numerous useful medical applications which we are currently striving toward," Irani said. "These include using our brain encoder/decoder to detect and analyze brain disorders such as MS, PTSD, autism and psychiatric disorders."
The scan constrains the picture; the mind remains its own.
via livescience.com/feeds/news (Source)
More from Priya Raman
Show full bio
Senior reporter covering people, launches and day-to-day practice in general science news at SciBeat.
378 articles
Nearby plates
- Scientists reconstruct 10-second videos from mouse brain signals
- The Brain's Fear Center Doubles as a Save Button for Memory
- £50m centre targets the mysteries of early brain development
- Alzheimer's Reshapes Brain Communication Before Symptoms Appear
- The Unconscious Brain Predicts Words, Baylor Study Finds