Efficient open foundation model for EEG reconstruction, denoising, and upsampling.
zyphra/ZUNA1.1
Copied
About
ZUNA1.1 is Zyphra's open, Apache 2.0-licensed EEG foundation model that reconstructs, denoises, and upsamples EEG recordings across arbitrary channel layouts and time segments. This 380M-parameter diffusion autoencoder conditions on the true 3D positions of each electrode, works on essentially any montage from 4-channel consumer headbands to 256-channel research caps without retraining and delivers high-fidelity reconstruction across variable-length real-world recordings. ZUNA1.1 can accept both a 0.5 s trial snippet and a 30 s continuous stretch, with no retraining or reconfiguration.
We delete electrodes from a specific brain region, and then reconstruct that region given the remaining seven regions, a more experimentally realistic setup than typical masked autoencoders which drop channels at random. This setup also mimics some consumer headsets which typically cluster electrodes primarily in one brain region. ZUNA1.1’s new training regime yields tangible benefits on more realistic scenarios for missing data.
Performance
ZUNA1.1 was trained on roughly 3.5 million channel-hours of public EEG using quality-aware preprocessing that scores every channel-second, letting it learn from partially-noisy recordings that conventional pipelines discard. Combined with training on four realistic channel-dropout patterns drawn from how EEG fails in practice, this produces consistently high-fidelity reconstruction across montages, sequence lengths, and noise conditions, outperforming classical interpolation methods.
