Brain-IT AI model recreates images from fMRI brain scans

Brain-IT turns fMRI brain scans into pictures. fMRI, or functional magnetic resonance imaging, tracks changes in blood oxygen in the brain, which rise when nerve cells in an area become active. Brain-IT reads those patterns and produces an estimate of the image the person was viewing.

Other systems have done this before, but they tend to get the gist right and the details wrong. According to Irani, they often slip on basics like composition and color. Brain-IT does better on both the content of the picture and its finer details, such as where objects sit and what color they are.

It also runs the other way. Show it a photo, and it predicts what a person’s brain scan would look like while viewing it. That two-way ability turned out to be the key to its speed.

Brain decoding has always had a data problem. Every brain responds a little differently, so a system trained on one person cannot simply be used on another. Gathering enough scans is slow, costly and uncomfortable for the volunteers lying in the scanner.

The Weizmann team worked with the Natural Scenes Dataset, a public collection from eight volunteers who each viewed thousands of photos across dozens of scanning sessions, more than 70,000 images in all. The researchers split each person’s scans into about 40,000 tiny three-dimensional units called voxels and tracked how each one responded to colors, positions, and things like faces or food.

Then they looked for patterns shared across all eight brains. Their encoder, the part that predicts brain activity from an image, sorted those responses into 128 “functional regions” that behave the same way in different people, even when they sit in slightly different places. Some were already known to neuroscientists. Others were new, including a split inside the brain area that handles places, with one part reacting to indoor scenes and another to outdoor ones.

Because the encoder can predict brain responses to any photo, the team fed it extra pictures no volunteer had ever seen. Those predicted scans became training material for the decoder, which turns brain activity back into images. In effect, the model built itself a much larger practice set without anyone spending more time in a scanner.

The result: with one hour of scans from a new person, Brain-IT matches what earlier methods achieved with about 40 hours.

Knowing someone saw a dog is not enough to draw the scene. You also need to know where the dog stood and what was around it. Brain-IT handles this in two parts.

One part creates a rough sketch of the layout, shapes and colors. Another supplies clues about the objects and what the scene contains. An AI image generator then turns that rough sketch into a detailed picture, guided by the content clues. During reconstruction, the system only receives the brain measurements, never the original photo.

Irani herself has called “mind-reading” a catchy label. Brain-IT does not capture thoughts, memories or language. It reconstructs what a person is looking at, based on the brain activity that the act of seeing produces.

It also has big practical limits. It only works with fMRI, which means lying still inside a large, expensive MRI machine, willingly, for long stretches. Nobody can scan your brain this way without you knowing. The researchers are looking at whether a simpler EEG headset, which records electrical signals through the scalp, could deliver similar results, but that has not been shown yet.

The nearest practical use is medical. The researchers say methods like Brain-IT could eventually help people with severe paralysis communicate, and give scientists a sharper tool for studying how the brain processes what we see.

The field is moving quickly. In 2017, a Purdue University model could predict the object a person was viewing, but its images were little more than blurs. In 2022, an Osaka University team used the Stable Diffusion image generator to get clearer results, though it needed text captions to help. Brain-IT gets closer to the real picture with far less data.

That pace is why privacy keeps coming up. The data comes from bulky hospital equipment for now, but as decoders improve and cheaper sensors are tested, questions about who can collect brain data, and what they can learn from it, will only get more pressing.

The team now wants to test whether the same approach works for sound. Dreams are the long-term goal, but they would require decoding video, and dozens of images flash past every second, far faster than an fMRI scanner can capture them. “If we overcome all these obstacles,” Irani said, reading dreams might one day be possible.

The work was presented at the 2026 International Conference on Learning Representations in Rio de Janeiro, and the team has published its code on GitHub.