Mapping fMRI research.

How can one voxel-level representation transfer across fMRI datasets?

Brain-DiT

ROI

How can a model learn across different brain states?

FlatClip

Surface

Can existing vision models provide useful features for fMRI?

SLIM-Brain

Which moments need expensive voxel-level computation?

FlexiBrain

How can one encoder handle heterogeneous acquisition resolutions?

BrainWorld

Can individual anatomy guide whole-brain fMRI dynamics?

DCA

fMRI Scaling StudyBrain-DiT

Can learned voxel features help discover individual brain parcels?

fMRI Scaling Study

ROI

Under a fixed budget, should we scale data, models or training time?

BrainTaskonomy

ROI

What should we pretrain on, and which tasks should we transfer?

Explore the 9 connections

ROI ↔ Voxel Research connection

Two complementary representation routes: Brain-DiT learns from atlas-defined ROI signals; Omni-fMRI learns directly from voxel volumes.

Omni-fMRIBrain-DiT

Surface ↔ Voxel Research connection

Two spatial interfaces: FlatClip reuses a vision encoder through cortical flatmaps; Omni-fMRI learns directly from voxel volumes.

Omni-fMRIFlatClip

ROI ↔ Surface Research connection

Different interfaces expose different spatial information: ROI time series for Brain-DiT and cortical flatmaps for FlatClip.

Brain-DiTFlatClip

Temporal efficiency Research connection

SLIM-Brain advances efficient voxel learning through time-window routing and sparse Hiera-JEPA computation.

Omni-fMRISLIM-Brain

Spatial / resolution Research connection

FlexiBrain extends voxel learning to native space and heterogeneous resolutions with Mamba-JEPA.

Omni-fMRIFlexiBrain

4D generation Research connection

From ROI-level conditional diffusion in Brain-DiT to anatomy-conditioned, whole-brain voxel dynamics in BrainWorld.

Brain-DiTBrainWorld

Atlas discovery Research connection

Voxel representation meets brain organization: DCA uses pretrained Swin-UNETR features and a spatial graph to discover individual parcels.

Omni-fMRIDCA

ROI scaling Research connection

Two ROI-level learning studies: Scaling controls the NeuroJEPA training budget; Brain-DiT learns representations through conditional diffusion.

Brain-DiT proxy Method reuse

Direct methodological reuse: BrainTaskonomy uses a lightweight Brain-DiT proxy to estimate domain difficulty and directed facilitation.

BrainTaskonomyBrain-DiT

Three questions to explore

01 / REPRESENT

How do we learn from brain activity?

Preserve spatial detail, adapt to acquisition differences, and make long recordings tractable.

Omni-fMRI FlexiBrain SLIM-Brain

02 / DISCOVER

What can those representations enable?

Explore conditional generation, whole-brain dynamics and individualized brain parcels.

Brain-DiT BrainWorld DCA

03 / TRAIN

How should we use data and compute?

Compare model reuse, training budgets and task-transfer strategies.

FlatClip fMRI Scaling Study BrainTaskonomy

Featured research

View all research

Generative modeling

BrainWorld

A structural prior brings anatomy and four-dimensional brain activity into one generative framework.

Explore BrainWorld

NeurIPS 2026 · Oral

Select a project

Explore the collection

Models & methods

Explore projects by modality
and research purpose.

/

9 research projectsBy manuscript version date

Accepted · NeurIPS 2026SurfaceRepresentation06 / 09Version · 2026-09-25

FlatClip

Turn cortical activity into flatmap sequences and reuse a frozen vision encoder for a lightweight surface-based learning baseline.

Cortical surfaceFrozen encoderSigLIP 2

fMRI Scaling Study — original research figure

Preprint · arXivROITraining & scaling05 / 09Version · 2026-09-23

fMRI Scaling Study

A controlled study of how data, model size and training duration affect generalization in fMRI foundation models.

Scaling lawsGeneralizationControlled evaluation

BrainTaskonomy — original research figure

Preprint · arXivROITraining & scaling04 / 09Version · 2026-09-09

BrainTaskonomy

Organize pretraining domains and downstream transfer with measured task relationships, curriculum learning and budget-aware source selection.

Curriculum learningTask transferTraining strategy

Published · MICCAI 2026 · Best Paper ShortlistROIGeneration / Representation02 / 09Version · 2026-06-21

Brain-DiT

Metadata-conditioned diffusion pretraining brings resting-state, task, naturalistic and other brain states into a shared representation framework.

Diffusion transformerMulti-stateMetadata conditioning

Accepted · NeurIPS 2026 · OralVoxelGeneration / Representation03 / 09Version · 2026-06-16

BrainWorld

Use individual brain anatomy as a structural prior to generate whole-brain fMRI dynamics across space and time.

4D generationStructural MRITemporal dynamics

Published · ECCV 2026VoxelRepresentation08 / 09Version · 2026-06-09

FlexiBrain

Learn from native-space fMRI across different resolutions with physical-space patching and a Mamba-JEPA encoder.

Native fMRIDynamic patch resizingMamba-JEPA

Accepted · ICML 2026VoxelRepresentation01 / 09Version · 2026-01-30

Omni-fMRI

Atlas-free brain representations with content-adaptive voxel patches. Preserve spatial detail while making large-scale pretraining more efficient.

Voxel-levelDynamic patchingMasked autoencoding

Preprint · arXivVoxelRepresentation07 / 09Version · 2026-01-30

SLIM-Brain

Route informative temporal windows to a hierarchical encoder to reduce the cost of voxel-level fMRI representation learning.

Sparse routingVoxel-levelEfficient learning

DCA — original Figure 2 showing pretraining and personalized atlas generation

Published · NeurIPS 2025VoxelAtlas09 / 09Version · 2025-09-20

DCA

Build personalized voxel-level brain atlases with pretrained fMRI embeddings and spatially guided clustering.

Brain parcellationGraph-guided clusteringNeurIPS 2025