AI Tool Reconstructs Images from Brain Scans with High Precision
- A new AI tool can guess what you’re looking at just by analyzing your brain scans and can recreate that image with remarkable precision.
- Neuroscientists have spent years attempting to reconstruct what people see from brain scans, but early efforts produced blurry, uninterpretable results.
- Unlike previous models that struggled to place objects correctly in generated frames, this brain decoder features two distinct branches.
A new AI tool can guess what you’re looking at just by analyzing your brain scans and can recreate that image with remarkable precision. Developed by Michal Irani and colleagues at the Weizmann Institute of Science in Rehovot, Israel, the system uses high-resolution functional magnetic resonance imaging data to decode brain activity and generate corresponding visual representations.
High Resolution Scans Improve Brain Decoder Accuracy
Neuroscientists have spent years attempting to reconstruct what people see from brain scans, but early efforts produced blurry, uninterpretable results. Irani and her research team bypassed traditional limitations by utilizing newer datasets gathered from scanners with a higher resolution where each voxel covers roughly one cubic millimeter of neurons. The team trained their model on data from eight individuals who viewed approximately 9,000 images inside an fMRI machine.
Unlike previous models that struggled to place objects correctly in generated frames, this brain decoder features two distinct branches. One branch predicts image structure while the other predicts content, feeding data into a diffusion model to refine the output. To scale up training without requiring endless human scanning sessions, the researchers paired their decoder with an encoder that predicts brain activity from images. This dual approach allowed the system to train on roughly 70 percent of images that were never originally shown to human participants in an fMRI scanner.
New Brain Encoder Reduces Calibration Time and Cost
Presented at the Cognitive Computational Neuroscience conference in New York, the resulting universal brain encoder requires minimal calibration for new subjects. While older decoding tools demand roughly 40 hours of fMRI data per individual—costing between $600 and $1,000 per hour—Irani’s decoder requires only one hour of data. None of us can afford 40 hours of imaging for a new subject,
said Tommy Sprague, a neuroscientist at the University of California Santa Barbara, noting that the efficiency could accelerate research speed.
Neurological Applications Raise Mental Privacy Concerns
Researchers see significant potential for neurological applications. Irani envisions the technology eventually helping locked-in patients communicate and allowing scientists to study the content of dreams or post-traumatic stress disorder flashbacks.
However, the technology raises serious ethical questions regarding mental privacy and consent. Sprague warned that surreptitiously extracting thoughts could turn decades of science fiction into reality. Marcello Ienca, a neuroscientist and philosopher at the Technical University of Munich, said that shifts toward EEG-based decoding could allow companies or courts to extract information without consent. Irani acknowledged potential misuse but noted she remains focused on positive applications.
Expanding Beyond Static Images
Although the current tool still experiences reconstruction failures—such as translating a cake into a pile of sandwiches or a dog in a bathtub into a goat—it outperforms existing alternatives by a significant margin. Irani is now working to move beyond static images into video and audio reconstruction. That’s something we don’t have yet, but we’re striving to achieve it,
she said.