Meta's TRIBE v2 Cuts Brain Scan Hours by 68% Using Synthetic fMRI Data
Meta AI researchers show that TRIBE v2's synthetic fMRI predictions can boost brain-to-image decoding accuracy by up to 68% in low-data regimes — but only when the real-to-synthetic ratio is carefully calibrated.
Takeaways
- Core finding: Meta AI researchers show TRIBE v2 synthetic fMRI augmentation improves brain-to-image decoding by up to 68% in low-data regimes. [Paper]
- How it works: TRIBE v2 generates synthetic fMRI responses to new images; these are mixed with real scans to train image decoders, with evaluation always on real fMRI only.
- BOLD5000 result: With just 30% of real data (~4 hours of scanning), augmentation matches full 13.3-hour baseline — saving ~9 hours per subject.
- Not plug-and-play: The optimal real-to-synthetic ratio varies by dataset and decoder type; too much synthetic data saturates or hurts performance.
- Zero-shot hint: Decoders trained exclusively on synthetic fMRI perform above chance in some settings, suggesting a path toward zero-shot brain decoding.
- Open source: TRIBE v2 model weights and code are publicly available under a non-commercial license. [Demo]
Brain-to-image decoding , the task of reconstructing what a person is looking at from their brain activity , has made remarkable strides in recent years. But it has a dirty secret: the best systems require tens of hours of fMRI scanning per subject, a constraint that puts cutting-edge decoding out of reach for most labs. A new paper from Meta AI researchers asks a pointed question: can you fill that data gap with synthetic brain scans generated by an AI?
Brain decoding is limited by the availability of labeled neural data and remains challenging in low-data regimes. To address this, the researchers investigate whether and when brain decoding can be boosted by augmenting small fMRI datasets with synthetic data generated by a pretrained model of fMRI responses to stimuli. The answer, laid out in their new paper, is a qualified yes , with important caveats about when it works and when it backfires.
The data wall blocking brain decoding
fMRI (functional magnetic resonance imaging) measures brain activity indirectly by tracking blood-oxygen levels across thousands of tiny 3D voxels. To train a decoder that maps these signals back to images, you need many paired examples of (image shown, brain response recorded). One of the most popular datasets for fMRI-to-image reconstruction is the Natural Scenes Dataset (NSD), which contains 30–40 hours of 7T fMRI per subject. Most neuroscience studies are far smaller and noisier , and modern decoders are strongest precisely in the data regime that few labs can afford.
Classical data augmentation tricks from computer vision , flipping, cropping, color jitter , don't translate to neural recordings. Neural data is inherently tied to the fixed biological architecture of the subject rather than the fluid, translation-invariant structure of a 2D image grid. So the Meta AI team takes a different approach: use a pretrained encoding model (a model that predicts brain responses from stimuli) to generate synthetic fMRI data for images the subject never actually saw.
TRIBE v2 as a synthetic brain data factory
Meta's FAIR team introduced TRIBE v2, a tri-modal foundation model designed to bridge the gap between AI and brain activity, predicting high-resolution fMRI responses across diverse naturalistic and experimental conditions. TRIBE v2 is a large encoding model pretrained on more than 1,000 hours of fMRI responses to video, audio, and language. Crucially, it was not trained to decode brain signals , it was trained to predict them from stimuli.