Brain-IT: Image Reconstruction from fMRI via Brain-Interaction Transformer
Paper • 2510.25976 • Published • 16
Error code: TooBigContentError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
This repository is a mirror of the NSD (Natural Scenes Dataset) data required by the Brain-IT project, for fMRI-to-image reconstruction research.
⚠️ For research use only. This repository is a data mirror only and does not own the data. Please comply with the usage licenses of each data source.
| Data | Source |
|---|---|
| Brain-IT original repository | https://github.com/WeizmannVision/brainit-fmri.git |
| Brain-IT paper | Brain-IT: Image Reconstruction from fMRI via Brain-Interaction Transformer (ICLR 2026, arXiv:2510.25976) |
| NSD dataset (fMRI betas / stimuli / experiment metadata) | Website https://naturalscenesdataset.org , public AWS S3 bucket https://natural-scenes-dataset.s3.amazonaws.com |
| NSD paper | Allen et al., A massive 7T fMRI dataset to bridge cognitive neuroscience and AI, Nature Neuroscience 2022 |
| COCO dataset (unlabeled images, not included in this repo) | Website https://cocodataset.org , unlabeled2017 download http://images.cocodataset.org/zips/unlabeled2017.zip |
| Algonauts ROI masks | Public ROI masks from the Algonauts Challenge (shared via Google Drive) |
brain-it-data/
├── experiment/ # NSD experiment metadata
│ ├── nsd_expdesign.mat
│ └── nsd_stim_info_merged.csv
├── imgs/
│ └── nsd_stimuli.hdf5 # Stimulus images (~73k images, ~37GB)
├── roi_masks/ # Algonauts ROI masks (sub1 ~ sub8)
│ ├── sub1/ ...
│ └── sub8/
└── raw_fmri/ # fMRI beta maps (subj01 ~ subj08, fsaverage space)
├── subj01/ ...
└── subj08/
.mgh files (betas_fithrf_GLMdenoise_RR, fsaverage space, ~468MB per lh/rh file).nsd_stimuli.hdf5 (stimulus images). The 123k COCO unlabeled2017 images are not included in this repository (they exceed the repo file-count limit); download them from the COCO website.data/scripts/run_all_downloads (options: --subjects, --inference-only, --skip-fmri, etc.).