High-Quality Image Reconstructions Achieved with Just 1 Hour of MRI Data

MindEye2 is a breakthrough method in fMRI-to-image reconstructions, cutting down training time from 40 to just 1 hour. Traditional methods require extensive fMRI data per subject, limiting practicality.

MindEye2 pretrains across 7 subjects, then fine-tunes on minimal data from a new subject, improving generalization. It maps brain data to a shared-subject space, then to CLIP image space, and finally to pixel space using Stable Diffusion XL.

By fine-tuning an unCLIP model, it achieves high fidelity reconstructions. MindEye2 outperforms other methods, enabling accurate reconstructions from a single MRI visit.

Notably, it achieves state-of-the-art performance with minimal training data, revolutionizing fMRI-based image reconstructions.

1 Like