Frequency-Decomposed Complex-State SSM Encoder for EEG Visual Decoding with Frozen Diffusion Evaluation
- Mapping EEG brain responses into CLIP image space with a frequency-aware Mamba-3 encoder.
EEG EncodingImage GenerativeSSMStable DiffusionMAMBA-3
About
A person looks at an image, an EEG cap records the brain response, and the goal is to turn that short, noisy signal into a 768-dimensional vector that lands in the same space as the CLIP embedding of the image they saw. A frozen Stable Diffusion 1.5 model then turns that vector back into a picture at evaluation time. The generator never trains, so any change in the results points back to the one thing we do train: the encoder. The proposed encoder splits the EEG into neural frequency bands and runs each band through a shared Mamba-3 state space stack, which is the contribution this project is built around.