Can fake brain scans teach us to read minds?

Researchers used TRIBE v2, a large model trained on 1000+ hours of fMRI data, to generate synthetic brain scans and mix them with real data to train image decoders. The synthetic data dramatically improved performance on two fMRI datasets, allowing the system to reconstruct which images people were viewing with far fewer real scans needed. This matters because fMRI scanning is expensive and time-consuming—synthetic augmentation could make brain-to-image decoding practical with minimal actual data collection.