New AI Recreates Images From Brain Scans With High Fidelity

Researchers at the Weizmann Institute built an AI decoder that predicts what a person is looking at from fMRI scans and recreates the image using a diffusion model. Training used scans from eight subjects at up to 1 mm³ neuron resolution, with an encoder generating synthetic scans to boost data.

• Decoder predicts image structure and content from brain scans
• Trained on scans from eight subjects shown thousands of images
• Encoder model generates synthetic scans to refine the decoder
• Requires expensive fMRI now, but could reach wearable EEG