• Checkpoints for Sub-1 and Sub2
• Generation results of SynBrain for all subjects
• Data for quick evaluation

Download files from HuggingFace SynBrain.

• Create conda environment using environment.yaml in the main directory by entering conda env create -f requirements.yml . It is an extensive environment and may include redundant libraries. You may also create environment by checking requirements yourself.

ps. You need to set your own path to run the code.

1. Download NSD data from NSD AWS Server;
2. Download "COCO_73k_annots_curated.npy" file from HuggingFace NSD;
3. Prepare visual stimuli and fMRI data;

   cd data
   python download_nsddata.py
   python prepare_nsddata_sclae.py -sub x
   

4. Extract CLIP image embedding by running extract_features_sdxl_unclip.ipynb

Run src/vae/train_vae.py for BrainVAE training

ps. Change sys.path/save_path/data_path to run the code correctly.

Run src/s2n/run_sit_os.sh for subject-specific S2N mapper training.

Run src/s2n/run_sit_os_ft.sh for subject-adaptive S2N mapper training.

Run src/s2n/generate.py for Visual-to-fMRI synthesis.

Run src/s2n/eval.py for Voxel-level and Semantic-level evaluation.

@article{mai2025synbrain,
  title={SynBrain: Enhancing Visual-to-fMRI Synthesis via Probabilistic Representation Learning},
  author={Mai, Weijian and Wu, Jiamin and Zhu, Yu and Yao, Zhouheng and Zhou, Dongzhan and Luo, Andrew F and Zheng, Qihao and Ouyang, Wanli and Song, Chunfeng},
  journal={arXiv preprint arXiv:2508.10298},
  year={2025}
}