AI & ML Practical Magic

We can now reconstruct the images you are seeing directly from your whole-brain activity.

April 14, 2026

Original Paper

MeshD²R: Mesh-guided Diffusion for fMRI-to-Image Reconstruction

SSRN · 6557878

The Takeaway

By using Mesh-guided Diffusion on whole-brain cortical topology, this framework creates high-fidelity visual reconstructions from fMRI signals. It moves neural decoding from category guessing to direct visual reproduction.

From the abstract

Current fMRI-to-image reconstruction methods often rely on manual region-of-interest (ROI) selection, potentially discarding informative signals and limiting decoding performance. We propose MeshD²R (Mesh-guided Dual-resolution Diffusion Reconstruction), a two-stage framework that leverages whole-brain cortical information. In the first stage, a Spherical CNN extracts global representations from cortical fMRI signals while preserving spatial topology. In the second stage, these representations a