EEG-to-Image Generation

Images generated by the original code from EEG recorded while a person looked at a daisy: rocks, a clock and shoes Before the fix: the person saw a daisy (left); the model drew rocks, a clock and shoes.

Images generated after the fix: every one is a daisy After the fix: daisies.

GWIT turns EEG brain signals into images with Stable Diffusion, but its published code did not reproduce the paper’s results.

At A*STAR’s Institute for Infocomm Research, I traced the fault to the ControlNet conditioning path and fixed its tensor-shape and initialization bugs.

Semantic accuracy (50-way top-1) rose from 0.02 to 0.94, on par with the paper’s 0.91.