Researchers built a system that reads brain activity and picks out which image a person was looking at - most of the time.
The team trained a compact neural network on THINGS-EEG2, a public dataset pairing EEG recordings with viewed images, using data from a single subject. Instead of trying to generate pictures from brain waves, the model mapped EEG signals to existing image embeddings and then had to pick the correct image out of a 200-image test set. It got the right answer in its top 10 guesses 58% of the time, far above the 5% chance baseline. The researchers also tried training generators to produce images directly from EEG without that embedding shortcut. Those attempts produced noise.
This matters because "brain-to-image" headlines tend to imply something closer to science fiction than what's actually happening in the lab. The real result here is narrower and more useful: EEG carries enough signal to narrow down what someone saw from a known set of options, not enough to paint a picture of it. That distinction - retrieval versus generation - is the difference between a plausible near-term tool and a much harder unsolved problem.
The model also failed to generalize. Applied to nine other subjects without retraining, performance collapsed, meaning whatever the network learned was tied closely to one person's brain, not brains in general. That's a bigger obstacle to any practical use than the pixel question. A system that needs to be retrained per user, on a lab-grade EEG cap, is a research result - not a product roadmap.