This paper introduces PictorialCortex, a novel framework for zero-shot cross-subject fMRI-to-image reconstruction, tackling the challenge of reconstructing visual experiences from brain activity of unseen individuals. It achieves this by disentangling stimulus-driven information from subject- and trial-specific variability using a compositional latent model and a new unified fMRI dataset.

This paper introduces Brain-Gen, a transformer-based framework that extracts spatio-temporal representations from EEG signals to reconstruct visual stimuli using Latent Diffusion Models. The method significantly improves clustering accuracy and generalization across unseen classes, advancing the semantic interpretation of EEG signals.