Researchers showed that a large vision transformer that has never seen EEG data can encode brain‑wave signals better than seven EEG‑specific foundation models. The trick is simple: instead of feeding raw numerical channels, the EEG is rendered as images – each channel becomes a row of picture patches. A frozen vision backbone (DINOv3 ViT‑B) is then paired with a tiny trainable display module (up to 1,113 parameters) and a readout head.

In a side‑by‑side test on eight clinical, sleep, affective and brain‑computer interface datasets, the frozen DINOv3 backbone (SeeEEG) topped all other models. Balanced‑accuracy ranks ranged from 1 to 7; the average rank was 1.3 versus 2.9 for the best EEG‑model baseline, REVE. Replacing the pretrained backbone with a randomly initialised ViT‑B dropped performance by 15.2 points on average, proving that the vision model’s visual pre‑training is key.

Beyond EEG data, the learned display layout transferred unchanged to a video encoder (V‑JEP 2.1), which outperformed every EEG foundation model on the same datasets.

The study purposefully held every encoder frozen, training only the display and readout heads. Cross‑validation on subjects was used to prevent shortcuts that could exploit subject‑specific idiosyncrasies, though the authors did not evaluate the method on entirely unseen acquisition setups. Results also come from eight carefully chosen datasets; performance on other clinical or BCI tasks remains unknown.