Brain-Diffuser

Official repository for the paper "Brain-Diffuser: Natural scene reconstruction from fMRI signals using generative latent diffusion" by Furkan Ozcelik and Rufin VanRullen.

Results

The following are a few reconstructions obtained :

Instructions

Requirements

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

Data Acquisition and Processing

1. Download NSD data from NSD AWS Server:

   cd data
   python download_nsddata.py

2. Download "COCO_73k_annots_curated.npy" file from HuggingFace NSD
3. Prepare NSD data for the Reconstruction Task:

   cd data
   python prepare_nsddata.py -sub 1
   python prepare_nsddata.py -sub 2
   python prepare_nsddata.py -sub 5
   python prepare_nsddata.py -sub 7

First Stage Reconstruction with VDVAE

1. Download pretrained VDVAE model files and put them in vdvae/model/ folder

   wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-log.jsonl
   wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-model.th
   wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-model-ema.th
   wget https://openaipublic.blob.core.windows.net/very-deep-vaes-assets/vdvae-assets-2/imagenet64-iter-1600000-opt.th

2. Extract VDVAE latent features of stimuli images for any subject 'x' using python scripts/vdvae_extract_features.py -sub x
3. Train regression models from fMRI to VDVAE latent features and save test predictions using python scripts/vdvae_regression.py -sub x
4. Reconstruct images from predicted test features using python scripts/vdvae_reconstruct_images.py -sub x

Second Stage Reconstruction with Versatile Diffusion

1. Download pretrained Versatile Diffusion model "vd-four-flow-v1-0-fp16-deprecated.pth", "kl-f8.pth" and "optimus-vae.pth" from HuggingFace and put them in versatile_diffusion/pretrained/ folder
2. Extract CLIP-Text features of captions for any subject 'x' using python scripts/cliptext_extract_features.py -sub x
3. Extract CLIP-Vision features of stimuli images for any subject 'x' using python scripts/clipvision_extract_features.py -sub x
4. Train regression models from fMRI to CLIP-Text features and save test predictions using python scripts/cliptext_regression.py -sub x
5. Train regression models from fMRI to CLIP-Vision features and save test predictions using python scripts/clipvision_regression.py -sub x
6. Reconstruct images from predicted test features using python scripts/versatilediffusion_reconstruct_images.py -sub x . This code is written as you are using two 12GB GPUs but you may edit according to your setup.

Quantitative Evaluation

Although results are expected to be similar, it may vary because of variations at reconstruction

1. Save test images to directory python scripts/save_test_images.py
2. Extract evaluation features for test images using python scripts/eval_extract_features.py -sub 0
3. Extract evaluation features for reconstructed images of any subject using python scripts/eval_extract_features.py -sub x
4. Obtain quantitative metric results for each subject usingpython scripts/evaluate_reconstruction.py -sub x

ROI Analysis

It has a bug that prevents to get the exact results but provides an approximation for most of ROIs, hopefully will be fixed soon.

1. Extract ROI fMRI activations for any subject 'x' using python scripts/roi_extract.py -sub x
2. Generate VDVAE, CLIP-Text, CLIP-Vision features forom synthetic fMRI using python scripts/roi_generate_features.py -sub x
3. Generate VDVAE reconstructions for ROIs using python scripts/roi_vdvae_reconstruct.py -sub x
4. Generate Versatile Diffusion reconstructions for ROIs using python scripts/roi_versatilediffusion_reconstruct.py -sub x

References

• Codes in vdvae directory are derived from openai/vdvae
• Codes in versatile_diffusion directory are derived from earlier version of SHI-Labs/Versatile-Diffusion
• Dataset used in the studies are obtained from Natural Scenes Dataset