Seeing Like a Brain: Predicting Subject-Specific EEG Responses to Natural Visual Stimuli

Predicting how natural visual stimuli evoke individualized neural dynamics is a fundamental challenge in cognitive neuroscience and brain–computer interfacing. Here, we present an open-source framework that directly predicts subject-specific, whole-head electroencephalography (EEG) responses from natural images. The model combines a visual encoder, EEG decoder and multi-scale dilated temporal refinement to generate 63-channel, time-resolved responses, with multi-objective training jointly constraining temporal, spectral and semantic consistency. Evaluation across 10 subjects and unseen test concepts demonstrates robust zero-shot prediction, with particularly strong fidelity over occipital regions. The predicted EEG preserves early visual components, spectral characteristics and image-related representational geometry while enabling millisecond-scale inference. Cross-dataset validation reproduces the main spatiotemporal, spectral and semantic trends in subjects with reliable stimulus information, and ablation analyses reveal complementary contributions of the model components and staged optimization. These results establish an efficient forward model linking natural images to individualized EEG responses and provide a scalable framework for studying visual neural dynamics and evaluating brain-like visual representations.