What is it about?

Can we reconstruct what a person has seen by measuring their brain activity? In this study, we introduce BOLD2Pix, a method that generates approximate reconstructions of viewed images from functional MRI (fMRI) recordings. We evaluate it using the Natural Scenes Dataset, in which participants viewed photographs while their brain activity was measured. Our approach first reduces the dimensionality of the fMRI data. Two compact neural decoders then estimate complementary information: what the image contains and how the scene is spatially arranged. A pretrained image generator combines these predictions to produce a reconstructed image.

Featured Image

Why is it important?

Reconstructing visual experiences from brain activity is an important challenge in neuroscience and artificial intelligence. Existing approaches often rely on large neural networks to translate complex brain measurements into representations that image generators can understand. BOLD2Pix explores a more compact alternative. By reducing the dimensionality of fMRI features and separately predicting semantic and spatial information, the method achieves reconstruction quality competitive with substantially larger approaches while using far fewer trainable parameters in its brain-decoding components. Our findings suggest that effective visual reconstruction does not necessarily require an equally large neural decoder, offering a direction for more efficient brain-to-image systems.

Perspectives

One of the most interesting questions raised by this work is how much information we actually need to extract from brain activity to reconstruct a visual experience. BOLD2Pix suggests that relatively compact representations of fMRI signals can provide useful semantic and spatial guidance to a powerful pretrained image generator. At the same time, current reconstructions remain imperfect, particularly when it comes to fine details and the precise arrangement of objects. Looking ahead, I believe an important direction is to improve reconstruction fidelity while exploring whether similar approaches can generalize beyond viewed images to imagined visual content and, eventually, to more accessible recording modalities such as EEG.

Ilia Semenkov

Read the Original

This page is a summary of: BOLD2Pix: A Compact Dual-Path Interface for fMRI-to-Image Reconstruction, November 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3840471.3842590.
You can read the full text:

Read

Contributors

The following have contributed to this page