MontageFlow and MontageDiff

Pretrained generators from MontageFlow: Physically Valid High Fidelity EEG Generation. Each model generates 16-channel bipolar-montage EEG windows at 200 Hz that satisfy the montage identities by construction. Code: https://github.com/sudarshanregmi/MontageFlow

Folder	Method	Corpus	Window	Labels
montageflow_tusz	MontageFlow (flow matching)	TUSZ	16 × 1000 (5 s)	0 background, 1 seizure
montageflow_tuev	MontageFlow (flow matching)	TUEV	16 × 1000 (5 s)	0 SPSW, 1 GPED, 2 PLED, 3 EYEM, 4 ARTF, 5 BCKG
montageflow_tuab	MontageFlow (flow matching)	TUAB	16 × 2000 (10 s)	0 normal, 1 abnormal
montagediff_tusz	MontageDiff (denoising diffusion)	TUSZ	16 × 1000 (5 s)	0 background, 1 seizure
montagediff_tuev	MontageDiff (denoising diffusion)	TUEV	16 × 1000 (5 s)	0 SPSW, 1 GPED, 2 PLED, 3 EYEM, 4 ARTF, 5 BCKG
montagediff_tuab	MontageDiff (denoising diffusion)	TUAB	16 × 2000 (10 s)	0 normal, 1 abnormal

Each folder holds config.yaml, model.safetensors (the generator, EMA weights) and prior.safetensors (the class-conditional latent prior).

Usage

git clone https://github.com/sudarshanregmi/MontageFlow.git && cd MontageFlow
pip install -r requirements.txt

import montageflow

model, prior, config = montageflow.load("sudarshanregmi/MontageFlow/montageflow_tuab")
x = montageflow.generate(model, labels=[0, 1, 1], prior=prior)
# x: (3, 16, 2000) float32, in microvolts / 100

MontageFlow samples with 50 Euler steps and MontageDiff with 50 DDIM steps.

Training data

The models were trained on the TUSZ, TUEV and TUAB corpora of the Temple University Hospital EEG Corpus, preprocessed as in JET (Wang et al., 2026). Access to these corpora requires registration with the Neural Engineering Data Consortium.

Citation

@misc{regmi2026montageflow,
  title  = {MontageFlow: Physically Valid High Fidelity {EEG} Generation},
  author = {Regmi, Sudarshan and Subedi, Nabaraj and Pillai, Arvind and Wu, Yu Yvonne and Chen, Yuliang
            and Heinz, Michael V. and Marsch, Lisa and Jacobson, Nicholas C. and Campbell, Andrew},
  year   = {2026}
}