A Review of Brain-Computer Interfaces Leveraging Generative AI for EEG Visual Reconstruction

Generative models are reshaping the research paradigm of visual reconstruction via Electroencephalogram (EEG) in Brain-Computer Interface (BCI) research. This article provides a systematic review of the cutting-edge advancements in this interdisciplinary field. Its primary contribution lies in revealing, from the perspective of model evolution, a paradigm shift in EEG-based visual reconstruction from discriminative classification towards generative reconstruction. Under the prevailing encoder-decoder architecture, the task framework has evolved from static images to encompass dynamic videos and three-dimensional models. Concurrently, generative models themselves have undergone iterative advancements from Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) to Diffusion Models, with each architecture presents distinct trade-offs in semantic consistency, structural fidelity, and training stability. Furthermore, this artick delineates the evolution of public datasets: from small-sample verification with early datasets like ImageNet-EEG, to large-scale datasets such as THINGS-EEG and Alljoined containing millions of trials, and further to newer benchmarks like SEED-DV and EEG-3D that feature temporal sequences and multimodality. While the scale of data has grown exponentially, bottlenecks including cross-subject consistency and suitability for generative tasks continue to challenge the field. Future breakthroughs necessitate concerted efforts in three key areas: the rational encoding of neurophysiological principles, the disentanglement of visual priors, and the establishment of standardized evaluation frameworks. Such advancements are crucial to transition EEG-based visual reconstruction from laboratory prototypes to real-world applications, thereby offering a novel computational modeling pathway for decoding the brain's visual representation mechanisms.