Reconstructions Gallery

The majority of the reconstructions capture the correct colors, camera viewpoint, and semantics but struggle to piece them together in a natural way (our Good column). We also see instances of poor reconstructions (Failure) and instances of faithful reconstructions (Success). These reconstructions are from a model trained on the training data from all subjects and tested on BOLD Moments Subject 01's testing set (the Ground Truth GIFs are sped up versions of the original 3 second videos). Note that a true reconstruction from brain activity may not be photorealistic as it depends on what the subject actually perceived and how stimulus features are represented throughout cortex. None of our reconstruction efforts employed any post-hoc selection procedure. To receive all reconstructions (including training set and baseline oracle reconstructions) please fill out this Google Form.