Visual Reconstruction from EEG

Decoding what someone sees from their brain activity, using deep learning.


This project reconstructs images from EEG brain signals. A subject views an image while wearing an EEG cap, and a two-pathway neural network decodes the brain activity into a visual reconstruction, all from less than a second of signal. Built on the THINGS-EEG2 dataset.

How it works

The system uses two complementary pathways to go from raw EEG to an image:

EEG Signal Visual Cortex Channels Semantic Pathway Diffusion Prior Reconstruction
  1. EEG Recording: Brain activity is captured from 63 electrodes across the scalp using a BrainVision actiCHamp device.
  2. Semantic Pathway: Maps EEG embeddings into CLIP embedding space, capturing high-level meaning (what the object is).
  3. Diffusion Prior: Translates the predicted embeddings into image space via Stable Diffusion, producing the final reconstruction.

Results

0.66
Avg. CLIP similarity
0.80
Best score
50
Test images
<1s
Signal window

Demo

Select a sample to watch the visual cortex respond in real time. The waveform shows 8 electrodes over the back of the head, where the brain processes vision. Two moments are marked:

Brain Activity — Visual Cortex — 0 ms
AI Reconstruction
AI reconstruction
Ground Truth
Ground truth
CLIP Similarity
0.000

Explore the Full Brain Response

The model uses all 63 EEG channels as input. Click any electrode on the head map to see that channel's recorded signal.

Visual Cortex
Parietal
Centro-Parietal
Central-Temporal
Frontal-Central
Frontal
Oz — Visual Cortex
800ms recording