acceptodds
Under review as a conference paper at ICLR 2027

MontageFlow: Physically Valid High Fidelity EEG Generation

Abstract

Generative modeling offers a promising solution to the scarcity of labeled clinical electroencephalography (EEG) data under strict privacy constraints. Although recent generative models report substantial improvements in generation fidelity, we show that these gains conceal a categorical failure. EEG is not an arbitrary multichannel signal: each channel records the voltage difference between two scalp electrodes. So, real recordings satisfy exact linear identities among their channels. Current generators invariably violate these identities, yet this has gone unnoticed. We therefore propose MontageFlow, a flow-matching model that enforces these identities by restricting its output to signals that satisfy them. To retain the EEG characteristics that oversmoothing washes out, we also condition the model on a physiological state vector encoded from the whole EEG window. For generation, this vector is drawn from a prior fitted in latent space. We find that ensuring validity enhances fidelity too. Across 3 large-scale benchmarks, our model improves TS-FID by 1.7–2.5× over the closest rival while ensuring physical validity.

Primary area:
learning on time series and dynamical systems
Keywords:
Generative modeling of EEG

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.25 over 4 reviews, above 96% of the others.
Reviewer_HNcgOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces MontageFlow, a flow-matching framework for physically valid, high-fidelity clinical EEG generation. The authors observe that bipolar EEG recordings satisfy exact linear identities: summing channels along two chains with shared endpoints must give identical signals, so the sixteen channels lie in a fourteen-dimensional subspace. They show that state-of-the-art generators violate these identities, yet existing metrics fail to detect this, and propose the montage index to quantify it. MontageFlow enforces validity by construction, projecting both the Gaussian source and the predicted velocity onto the constraint matrix's null space via a closed-form projector; the decoder is a 1-D DiT-B transformer conditioned on timestep, class label, and a window-level latent physiological vector. This latent, produced by a causal-Transformer encoder and sampled at generation from a class-conditional rectified-flow prior, preserves low-energy transients and heavy-tailed amplitudes that oversmoothing erases. Across TUSZ, TUEV, and TUAB, MontageFlow improves TS-FID by 1.7–2.5× over the strongest baseline while achieving a montage index of exactly zero; ablations show that latent conditioning and training through the projector each contribute substantial gains.

Critical strengths

++This work has a certain degree of innovation and is relatively interesting. The corresponding experimental results also verify the effectiveness of the algorithm.

++This paper is well written and easy to follow.

++The ablation experiments in the experimental section also verify the effectiveness of each design. The analysis of experimental results corresponds to the motivation in the introduction section.

Critical weaknesses

-- I recommend that the authors add a discussion of the model's limitations and directions for future research.

-- Although the overall analysis in the experimental section responds to the motivation, I still recommend that the authors conduct further detailed analysis on the content of the ablation part, so that scholars can better understand the relevant details.

Questions for authors

See Weakness

2 comments
AuthorsOpenReview ↗

Limitations and future work

We thank Reviewer HNcg for finding the core observation "original and important" and the projector "elegant and parameter-free." We also appreciate the reviewer’s recognition that validity and fidelity are complementary rather than competing.

Limitations and future work. We will add this section to the paper:

  • Scope of the guarantee. The identities hold for the bipolar view. They are necessary, not sufficient for physiological realism. So, we report them together with fidelity and diagnostics. A common-reference montage has no such identities, and the method reduces to its unconstrained form.
  • Future work. Extension to other biosignals with linear constraints such as ECG would be interesting future work.

Furthermore, we will revise the paper with the new empirical results, ablations and analyses run during the rebuttal. We thank the reviewer again for supporting this work.

I will ‌maintain my current score‌ for this submission.

Reviewer_k8hwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MontageFlow, a novel flow-matching model that takes into account the fact that EEG signals are voltages between two electrode points, leading to stronger EEG generation with important EEG characteristics.

Critical strengths

The paper is well-written. The motivation of the methodology is well-grounded on the montage geometric characteristics and constraints. The authors provide a very good preliminary section. The authors also provide multiple experimental results and ablation studies to support their claims as well as visualizations that show the generation of EEG signals.

Critical weaknesses

The comparisons are only limited to a specific dataset family. Therefore, generalization claims cannot be made sufficiently

According to the results section, much of the result improvement come from latent conditioning which weakens the claim for the necessity of montage constraint.

Critical reconstruction comparison with current SOTA EEG tokenizers (that are used in EEG FMs) is missing .

Minor: Maybe the related work could be after the Introduction section for better readability.

Questions for authors

Could the authors perform some analysis of different datasets that use different devices and montages?

Is latent conditioning the most critical part? If so, any ablation in other methodologies to validate it would be useful.

Can the authors provide a comparison with EEG tokenizers (e.g. codebook-based) to show the benefit of this method?

7 comments
AuthorsOpenReview ↗

Part 1/2: New corpora, montages, and what projection adds

We thank Reviewer k8hw for acknowledging the well-grounded motivation and preliminaries based on montage geometric characteristics and constraints, along with the multiple experimental results and ablation studies.

Generalization

We thank Reviewer k8hw for the opportunity to showcase our method's empirical results in additional datasets. Hence, we report the results (in terms of TS-FID, diagnostics metrics) obtained on four more corpora at 100k training steps at 10s window.

Test TS-FID :-

CorpusVanilla diffusionVanilla flow matchingJETJET×Latent, no projectionMontageFlowProjection improvement
CHB-MIT171.0231.3236.4248.2122.277.537%
NMT57.078.1182.1180.898.741.559%
Siena186.1248.8292.4287.8149.1132.111%
Helsinki177.0268.5205.0190.5119.281.831%

Diagnostics (average distance to real data, lower is better) :-

CorpusVanilla diffusionVanilla flow matchingJETJET×Latent, no projectionMontageFlowProjection improvement
CHB-MIT0.8200.9790.5210.4200.3900.26532%
NMT0.7721.2390.5850.5240.7720.37651%
Siena0.8830.6520.6590.6690.2000.17413%
Helsinki1.3361.9980.6940.3590.4130.31324%

MontageFlow has the lowest TS-FID and diagnostics on these corpora. Projection lowers TS-FID non-trivially in all cases. (Note: NMT with no projection was re-run with different seed since first run diverged. Against second run, MontageFlow still outperformed by 30% in TS-FID.)


Common-average montage

We also perform experiments on TUH corpora on common-average referenced EEG using the projector I−1C11⊤I - \tfrac{1}{C}\mathbf{1}\mathbf{1}^\top. We present the results below:

ProjectionTS-FIDDiagnostics (average)validity (0=perfectly valid)
none101.7 / 107.4 / 100.70.634 / 0.243 / 0.2140.74 / 1.36 / 0.28
MontageFlow66.0 / 100.4 / 75.70.352 / 0.208 / 0.2410

It shows that MontageFlow is extendable to other montage such as common-average referenced EEG too.

Window length. We also vary along the window length in bipolar montage setup and find that the projection's gain grows with longer windows. The result at ~150k iteration is shown below:

5 s windows10 s windows
TUSZ−24%−27%
TUEV−8%−12%
TUAB−39%−59%
CHB-MIT−7%−12%

"much of the result improvement come from latent conditioning which weakens the claim for the necessity of montage constraint."

  • TL;DR: In raw TS-FID points, yes. But each step starts from a better model, so later steps have less left to gain. Measured by the share of the remaining gap to real data, projection closes about as much as latent conditioning (41% vs 43% on average). It is also the only source of validity.
  • A fair comparison. For each step of the ablation ladder (Table 4 of the revised PDF), we take its drop in test TS-FID as a share of the gap to real data that was still left (real vs. real: 39.1 / 25.3 / 23.3).
StepTS-FID change (TUSZ / TUEV / TUAB)Remaining gap closedAverage
+ latent conditioning−58 / −90 / −7147 % / 43 % / 40%43%
+ projection−23 / −10 / −7742 % / 10 % / 71%41%

Such relative contribution of latent conditioning and projection are comparable.

  • Validity is a result in itself. Every sample of all five baselines fails the identity check, and so does every sample of our own latent model trained without projection. No MontageFlow sample fails validity check.

  • The latent is ours too. Latent conditioning is established in image generation. For EEG, we identify the failure it fixes: oversmoothing of EEG signals. We provide the design it needs: one state vector per window, perturbation during training and learned prior to make generation possible.

  • What the constraint adds beyond validity, with the latent held fixed:

    1. TS-FID falls by 27% / 12% / 59% at 10 s windows.
    2. Diagnostic metrics The average distance to real data of the five diagnostics of Section 6 is reduced non-trivially.
    3. Tails and transients: The average distance to real data in terms of five tail and transient metrics (kurtosis, peak amplitude, crest factor, line length, transient rate) is also reduced.
Average distance to real dataTUSZTUEVTUAB
Diagnostics, without → with projection0.432 → 0.343 (−20%)0.240 → 0.221 (−8%)0.268 → 0.118 (−56%)
Tails and transients, without → with projection0.919 → 0.889 (−3%)0.657 → 0.526 (−20%)0.655 → 0.519 (−21%)

AuthorsOpenReview ↗

Part 2/2: Latent transfer, EEG tokenizers, related work

"Is latent conditioning the most critical part? Any ablation in other methodologies?"

  • Latent-conditioning transfers to other methods too. We added our latent to JET× (without its auxiliary losses). We present results in TS-FID, and montage index below:
TUSZTUEVTUABMontage index
JET× (no auxiliary losses)195.3342.9256.10.334 / 0.404 / 0.396
JET× (no auxiliary losses) + latent-conditioning136.9125.6172.00.289 / 0.233 / 0.224

Our latent conditioning lowers JET×'s TS-FID by 58 / 217 / 84, so it is a general tool. It does not give validity though.


Comparison with EEG tokenizers. We compare MontageFlow with LaBraM's released tokenizer in 10s window. Both encode the same 2,000 test windows and decode them back to EEG. The results (test TS-FID) are presented in the table below:

ModelTUSZTUEVTUAB
MontageFlow54.1173.954.8
LaBraM tokenizer226.9301.8339.9

MontageFlow performs substantially better than the EEG tokenizer (LaBraM) on all three corpora, including TUSZ, which is part of LaBraM's pre-training data. Even MontageFlow's samples from the learned prior, which never see the test windows, reach 52.7 / 95.3 / 54.7.


Related work placement. Thank you for the suggestion. We'll take it into account. We kept the section order in this revision so that section references in the reviews stay valid.


We hope these additions address the reviewer's concerns. We are more than happy to discuss further. Please let us know if there is any outstanding concern left.

AuthorsOpenReview ↗

Please let us know if the concerns have been addressed.

Thank you again for the valuable suggestions and comments. We appreciate it.

We would like to sincerely request to let us know if there are any more concerns left to be addressed. We highly appreciate your time and engagement with our work.

Reviewer_WNxYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

(1) This paper proposes MontageFlow, a flow-matching framework for EEG generation, motivated by the need for more labeled clinical EEG data.

(2) The authors combine flow matching with a fixed montage projector and learned window-level conditioning to generate samples that satisfy the montage identities.

(3) Experiments on TUSZ, TUEV, and TUAB show that the generated EEG satisfies these identities: two chains with the same endpoints sum to the same value at every timepoint.

Critical strengths

(1) The paper is well organized and easy to follow.

(2) The experimental results show the potential of the framework to generate EEG for data augmentation.

Critical weaknesses

Motivation:

(1) The authors state "We therefore enforce these identities by construction rather than by penalizing violations" in Section 1. Generating signals with a common reference and then applying a fixed bipolar difference layer also guarantees these identities. Does MontageFlow offer better generation quality or training efficiency than this alternative?

(2) The authors describe "synthesizing data where labeled recordings are scarce or cannot be shared" in Section 1. However, no experiment tests whether adding generated EEG improves classification or detection on a held-out real test set. The intended use of the generated data and its practical benefit remain unclear.

Related Work:

(1) The authors describe "a linear equality with a fixed, closed-form projector" in Section 7. However, the review omits recent work on constrained flow matching, such as Physics-Constrained Flow Matching [1]. The authors should clarify what is new in their flow-matching design for EEG.

(2) The authors state "Per-channel augmentation may also violate montage identities" in Section 7. However, the review omits relevant EEG generation methods, such as the diffusion model of Vetter et al. [2]. The authors should specify which methods violate the identities and why. Could augmentation of the original reference-channel signals, followed by bipolar conversion, avoid this problem?

[1] Utkarsh U, Cai P, Edelman A, Gomez-Bombarelli R, Rackauckas C V. Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints. Advances in Neural Information Processing Systems, 2025.

[2] Vetter J, Macke J H, Gao R. Generating realistic neurophysiological time series with denoising diffusion probabilistic models. Patterns, 2024, 5(9): 101047.

Method:

(1) The authors state "the velocity network's output is always projected onto ker A" in Section 2.2. This restriction uses a standard orthogonal projection. Together with conditional flow matching and a learned latent prior, the method appears to be an incremental combination of existing techniques. The authors should identify the technical contribution beyond applying these components to the chosen EEG montage.

(2) The authors state "MontageFlow uses unprojected Gaussian noise during training but projects the initial noise at sampling time" in Section 5. Training inputs generally violate the montage identities, whereas sampling remains within the valid subspace. The projected velocity cannot remove the violations along the specified training path. I wonder how the learned velocity transfers to the sampling trajectory. The noise ablation supports this empirical choice, but does not explain the mismatch between training and sampling.

Experiments:

(1) The authors state "All three are preprocessed with the released pipeline of Wang et al. (2026)" in Section 3. The paper reports window counts but does not clearly specify the data partition procedure or whether subjects are disjoint across training and test sets. It is therefore unclear whether the evaluation tests generalization to unseen subjects.

(2) The authors state "Conditioning both the prior and the velocity network on y preserves the dependence between latent and label seen in training" in Section 2.2. However, the aggregate generation metrics do not establish whether samples match their target classes. Class-specific fidelity and coverage are not reported, which is particularly important for the imbalanced TUEV dataset.

(3) The authors state "We observe that latent conditioning mitigates oversmoothing of EEG" in Section 1. The ablation shows improvements in amplitude and TS-FID, but does not explain which information the latent encodes or how that information reduces oversmoothing. Could the authors provide more insight into this mechanism?

(4) The authors state "Capturing dominant low-frequency activity while preserving low-energy transients and heavy-tailed amplitudes remains challenging" in Section 1. However, it is unclear how montage projection helps preserve these features. Table 4 measures its effect on aggregate TS-FID, while the physiological diagnostics do not isolate its contribution from latent conditioning. A controlled comparison of transient morphology and amplitude tails would help establish this link.

Questions for authors

Please see the Weaknesses section.

6 comments

We thank Reviewer WNxY for finding the paper well organized and for noting the potential of the framework. We sincerely thank the reviewer for this opportunity to answer precise insightful questions, which has genuinely improved the paper.


Generating referential signals and applying bipolar difference: Does MontageFlow offer better generation quality or training efficiency than this alternative?

Yes, precisely. MontageFlow offers better generation quality. We present the comparison in terms of TS-FID and diagnostics metrics in the following table:

MethodTS-FIDDiagnostics (average)
MontageFlow66.0 / 115.7 / 54.70.334 / 0.221 / 0.118
Referential, then bipolar difference102.9 / 125.2 / 118.40.504 / 0.336 / 0.220

The results clearly show that the MontageFlow offers better generation quality.


Does generated EEG help on a held-out real test set? For every method, we train CBraMod classifier on generated data only and test on the real test split. We reports the results in terms of AUROC or kappa below:

Training setTUSZ (AUROC)TUEV (kappa)TUAB (AUROC)
EEG-GAN0.4990.0520.493
Vanilla diffusion0.8520.3850.794
Vanilla flow matching0.7300.2430.692
JET0.8290.2990.728
JET×0.7660.1290.673
MontageFlow0.8820.4430.816

The results show that MontageFlow gives the highest AUROC or kappa. Furthermore, where labels are scarce (1% of the real training windows) on imbalanced dataset (TUEV), adding 5,000 MontageFlow windows per class raises balanced accuracy from 0.387 to 0.565.


PCFM. We will cite PCFM. It enforces constraints at inference on a pretrained flow. We train through a closed-form projector. We'll make the difference between PCFM and MontageFlow more clear in the related works section. We applied PCFM to our latent model trained without projection and compare with MontageFlow at 100k training steps and 10 s windows:

CorpusTS-FID, PCFMTS-FID, MontageFlowDiagnostics, PCFMDiagnostics, MontageFlow
CHB-MIT338.277.50.9300.265
NMT47.641.50.9580.376
Siena203.7132.10.5090.174
Helsinki108.681.80.6940.313

MontageFlow has the lower TS-FID, and is substantially closer to real data wrt. PCFM on the diagnostics metrics everywhere.


Vetter et al. and per-channel augmentation. We will cite Vetter et al. (2024) in the upcoming revision. You are right about augmentation: augmentation can be easily followed by necessary subtraction operation to make channels montage-consistent.


What is new?

ContributionEvidence
Problem: bipolar generators violate the montage identities, and TS-FID does not see itEvery sample of every baseline fails the bipolar montage identity check
Fix: train through a closed-form projectorPerfect montage consistency, brings generated data closer to real EEG
Latent conditioning for EEG: fixes oversmoothingAmplitude 42–52% → 78–84% of real, test TS-FID −58 / −90 / −71

We will revise PDF to make it even more clear.


How does the velocity learned on off-subspace states transfer to sampling? Thank you for this question. We have improved the sampling as mentioned in common response. Now, we sample exactly as we train: the off-subspace noise decays as (1−t)Qε(1-t)Q\varepsilon and vanishes at t=1t = 1 hence samples stay exactly valid. The revised PDF already uses improved sampler.


Are subjects disjoint between training and test? Yes. We use the official TUH train/eval partitions. No subject ID appears in both for TUAB or TUSZ. TUEV's evaluation IDs are anonymized. Its documentation states that the sets are disjoint, and no window appears in both.


Class-specific fidelity and coverage. Per-class test TS-FID and recall averaged over the classes of each corpus is reported in the Table below:

Class averageTS-FID, MontageFlowTS-FID, best baselineRecall, MontageFlowRecall, best baseline
TUSZ73.2144.9 (diffusion)0.510.31 (flow matching)
TUEV244.7271.7 (diffusion)0.180.18 (flow matching)
TUAB58.6164.6 (diffusion)0.380.15 (diffusion)

MontageFlow has the lowest class-averaged TS-FID on all three corpora and the highest recall on TUSZ and TUAB. On the imbalanced TUEV, MontageFlow ties with vanilla flow matching (best baseline).


Do samples match their target class? We perform an experiment for this: We train a CBraMod classifier trained on real data on TUH corpora. Then, we classify 5,000 generated windows per class and obtain (balanced accuracy / worst-class recall) which is reported below:

Windows scoredTUSZTUEVTUABAverage
EEG-GAN0.466 / 0.0220.281 / 0.0040.500 / 0.0000.416 / 0.009
Vanilla diffusion0.670 / 0.3480.813 / 0.1520.741 / 0.7130.741 / 0.404
Vanilla flow matching0.636 / 0.2740.843 / 0.2860.798 / 0.7030.759 / 0.421
JET0.750 / 0.6900.341 / 0.0550.526 / 0.0540.539 / 0.266
JET×0.754 / 0.5540.930 / 0.6050.808 / 0.6660.831 / 0.608
MontageFlow0.796 / 0.6440.937 / 0.7890.799 / 0.7170.844 / 0.717

MontageFlow achieves the highest average balanced accuracy and worst-class recall. It indicates the strongest overall agreement between generated samples and their target labels under the real-data-trained classifiers.


What does the latent encode? A linear probe, fit on training windows, recovers simple properties of each test window from its latent alone. We report the share of each property's variation across test windows that the probe explains (1 = fully recovered, 0 = not at all):

PropertyTUSZTUEVTUAB
Window amplitude0.770.640.98
Delta power0.810.750.96
Spectral slope0.560.010.85

The latent carries each window's amplitude and slow-wave (delta) power on all three corpora. Finer spectral shape is recovered on TUSZ and TUAB, but not on TUEV.

Swapping only the latent (same noise and label) between the 100 quietest and 100 loudest test windows of a class moves the generated amplitude by 69–83% / 29–83% / 87–93% of the real gap. The latent tells the decoder which window it is producing, so the objective loss function no longer pulls the output toward an average.


Does projection itself help transients and tails? Yes. We report metrics for tails and transients along with diagnostics metrics in the format of without → with projection, with the latent held fixed, across TUH corpora in the Table below:

Average distance to real dataTUSZTUEVTUAB
Diagnostics metrics0.432 → 0.343 (−20%)0.240 → 0.221 (−8%)0.268 → 0.118 (−56%)
Tails and transients0.919 → 0.889 (−3%)0.657 → 0.526 (−20%)0.655 → 0.519 (−21%)

Tails and transients metrics is the average distance to real data in terms of following metrics: kurtosis, peak amplitude, crest factor, line length, and transient rate. It shows montage projection being consistently helpful.


Thank you again for the wonderful review. We hope these additions address the reviewer's concerns. We will add all new results to the paper. We are more than happy to discuss further. Please let us know if there is any outstanding concern left.

Thank you again for the in-depth review with valuable suggestions and comments. We appreciate it.

We would like to kindly request to let us know if there are any more concerns left to be addressed. We highly appreciate your time and engagement with our work.

Reviewer_Ly8oOpenReview ↗
Rating
4(+3)
Confidence
2
Informativeness
2(+1)
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The authors build a model to generate synthetic EEG. The motivation and premise land on the notion that each EEG channel records the voltage difference between two scalp electrodes and so real recordings must satisfy arithmetic relations between channels, and that existing generators break these assumptions. Their model forces generated signals to obey them and reports better similarity to real EEG. However, this premise is flawed; the truth is closer to the opposite. EEG is recorded using a common references and signals from the electrodes are related but independent. Bipolar montages such as the one used here are from offline re-referencing, which is also rather unusual in clinical and neuroscience EEG applications. I thus do not find the results compelling. The authors' own analysis shows most of the gain comes from another part of the model not from enforcing the montage relationships.

After all the revisions the authors have made, I will accept the paper.

Critical strengths

The evaluation includes physiologically meaningful checks, such as power spectrum, alpha peak, amplitude over time, distribution etc., not a single summary score.

Critical weaknesses

the fundamental premise is flawed; the truth is closer to the opposite. EEG is recorded using a common references and signals from the electrodes are related but independent. Bipolar montages such as the one used here are from offline re-referencing, which is also rather unusual in clinical and neuroscience EEG applications. I thus do not find the results compelling. the authors' own analysis shows most of the gain comes from another part of the model not from enforcing the montage relationships.

Questions for authors

Have the authors checked the raw data from the Temple University datasets?

2 comments
AuthorsOpenReview ↗

Part 1/2: The premise holds for any reference; one wording in abstract corrected

We thank Reviewer Ly8o for the careful reading, and noting our thorough evaluation.

"EEG is recorded using a common references ... Bipolar montages such as the one used here are from offline re-referencing"

TL;DR: Agreed: EEG is recorded against a common reference, and our method never assumes otherwise. One sentence of our abstract blurred recording and montage. That is just one wording error, not a flaw in the method. We fixed it, and no result changes.

  • Abstract. Before: "each channel records the voltage difference between two scalp electrodes." Now: "in a bipolar montage, each channel is the voltage difference between two scalp electrodes." Section 1 already said it correctly ("acquired from scalp electrodes, but read as potential differences between pairs of them").
  • We would also want to correct the reviewer's summary from "premise land on the notion that each EEG channel records the voltage difference" to "premise land on notion that in bipolar montage, each EEG channel is read as voltage difference"

"the fundamental premise is flawed; the truth is closer to the opposite."

Our premise is the bipolar montage, whose channels are differences of electrodes and so satisfy exact identities however the EEG was recorded. The paper generates EEG in the bipolar montage, a standard clinical layout. Subtraction cancels the reference, so the identities are exact for any recording. We have revised to narrow the wording to the bipolar montage throughout.

"[Bipolar montages] ... are also rather unusual in clinical ... EEG applications."

  • TL;DR: No, it is a standard clinical montage, these corpora are labelled on it. Recent EEG models and generators use it.
  • The corpora themselves. TUSZ seizure annotations are stored per bipolar channel, and TUEV event labels index TCP channels. The TUH documentation calls this "a standard ACNS TCP montage" and "our preferred way of viewing seizure data." CHB-MIT, recorded at Children's Hospital Boston, is distributed as bipolar channels.
  • Guidelines. ACNS Guideline 3 (Acharya et al., 2016) lists longitudinal bipolar as one of three standard montage classes, and recommends using both bipolar and referential montages for interpretation.
  • Common in deep learning on EEG. JET (ICML 2026), BIOT (NeurIPS 2023), CBraMod (ICLR 2025), CodeBrain and TFM-Tokenizer (ICLR 2026), CSBrain, REVE, LUNA and NeurIPT (NeurIPS 2025), EEG2Rep (KDD 2024) and Gijsen & Ritter (ICML 2025) use bipolar montages as input. So do expert-level clinical detectors such as SpikeNet (JAMA Neurology 2020) and SPaRCNet (Neurology 2023). Based on these evidences, we respectfully disagree on bipolar montage being unusual in clinical EEG applications. However, we acknowledge that referential and average-reference data are more common in cognitive neuroscience. Our scope is about clinical bipolar EEG and the revision says so. Our work is extensible to average-reference data too. Results in later section.

"most of the gain comes from another part of the model not from enforcing the montage relationships."

  • TL;DR: True for TS-FID alone, but TS-FID cannot see validity, that other part is also our contribution, and with it held fixed the projection still improves fidelity on every corpus.
  • Validity is a result in itself. TS-FID is computed from FFT magnitudes and cannot see the identities, so it should be read together with the montage index. On validity: every sample from all five baselines fails the identity check, and no MontageFlow sample does. Why validity matters in practice is shown in the later section.
  • The latent is our contribution too. Latent conditioning is established in image generation, but not in EEG generation. We identify the failure it fixes in EEG: generators trained with a regression loss oversmooth, keeping only 42–52% of the real amplitude. JET targets waveform fidelity with hand-specified auxiliary losses whose weights need per-dataset tuning. A learned latent instead restores 78–84% of the amplitude.
  • The projection improves fidelity. Please see the effect of projection in one common setting (10 s windows) in corresponding test split (format: no projection -> projection).
CorpusTS-FIDDiagnostics, averageMontage index
TUSZ72.2 → 52.7 (−27%)0.476 → 0.4390.360 → 0
TUEV108.2 → 95.3 (−12%)0.261 → 0.2120.285 → 0
TUAB131.8 → 54.7 (−59%)0.268 → 0.1180.224 → 0
CHB-MIT71.5 → 63.1 (−12%)0.224 → 0.2110.137 → 0
NMT98.7 → 41.5 (−58%)0.772 → 0.3760.345 → 0

(NMT is trained for 100k steps.) We hope a 12–59% TS-FID reduction is not discounted because the latent's gain is larger. Also, please see diagnostics (average) and montage index too.

AuthorsOpenReview ↗

Part 2/2: Why validity matters, and generating in other montages

Why validity matters in practice

  • TL;DR: Only valid samples can be viewed in another montage.
  • Why. Clinicians read the same recording in several montages, such as bipolar and average reference. Consider 16 channel electrode and per-hemisphere average reference for ease of explanation. Converting from bipolar to an average reference first rebuilds each electrode by adding bipolar channels along a chain. O1 has two chains: temporal (Fp1-F7-T3-T5-O1) and parasagittal (Fp1-F3-C3-P3-O1). A valid sample gives the same O1 either way. An invalid one gives two different O1 signals, so its average-reference view is ill-defined.
  • New experiment. We converted generated windows to the average reference through each chain and measured the difference between the two results, as % of the signal's RMS (TUSZ / TUEV / TUAB):
GeneratorDifference
MontageFlow0 / 0 / 0
Vanilla diffusion11 / 18 / 10%
JET×15 / 28 / 14%
Vanilla flow matching28 / 35 / 18%
EEG-GAN55 / 49 / 53%
JET81 / 81 / 81%
MontageFlow without projection63 / 70 / 37%

Only MontageFlow samples have a single average-reference view. The last row shows this comes from the projection, not the latent. Even in that view MontageFlow also has the lowest TS-FID on every corpus (58.9 / 100.1 / 54.8; best baseline 105.9 / 128.6 / 119.2).

"Have the authors checked the raw data from the Temple University datasets?"

Yes. TUH provides raw files too in their recorded form, hence we demonstrate more results below in other montages.

  • Referential and common average generation, then bipolar conversion. We trained our latent-conditioned generator on referential EEG and on common-average EEG built from raw files with no constraint and derived the bipolar channels from them. Even these results favour MontageFlow:
MethodTS-FIDDiagnostics, averageMontage index
MontageFlow66.0 / 115.7 / 54.70.334 / 0.221 / 0.1180
Referential, then subtract102.9 / 125.2 / 118.40.504 / 0.336 / 0.2200
Common average, then subtract149.5 / 139.6 / 121.70.750 / 0.340 / 0.1970

What if average-referenced were the target montage?

Suppose bipolar EEG were set aside and users needed average-referenced EEG. MontageFlow would still be the better route, in two ways:

  • Generate in common average directly, with our projector. Common average has one identity (channels sum to zero), and the method applies to any montage with linear identities. Training through the zero-sum projector lowers TS-FID by 35% / 7% / 25% (TUSZ / TUEV / TUAB) against the same generator without it.

  • Generate in bipolar, then convert. MontageFlow's samples are valid, so they convert to a single average-reference version (each hemisphere to its own average as the 16 chains do not link the hemispheres). We compare against our latent-conditioned model trained on common average, converted the same way (TUSZ / TUEV / TUAB):

GeneratorTS-FID, average reference
MontageFlow (bipolar), converted58.9 / 100.1 / 54.8
Latent-conditioned model trained on common average, converted105.2 / 108.5 / 107.0

Converting MontageFlow's bipolar samples gives a lower TS-FID than training on common average directly.


We hope these results and the corrected wording address the reviewer's concerns. We warmly welcome further discussions if there is any outstanding concern left.

Comments on the paper

5 comments
AuthorsOpenReview ↗

Note on all responses: MontageFlow is now sampled as it is trained (results further improved)

Minor update to sampling:

We now sample MontageFlow exactly as it is trained. Sampling starts from full Gaussian noise and the noise component outside the valid subspace decays to zero along the path. (The submission started from (projected) noise that already satisfied the identities.)

The results improved everywhere. The revised PDF will be uploaded as soon as possible. [Same checkpoints; generated samples still remain physically valid.]

AuthorsOpenReview ↗

All concerns addressed and reframed.

Thank you for taking the time to follow up, and for explaining the point so clearly. You are right, and we have followed your recommendation.

What we changed in the paper

  • Title: "MontageFlow: Montage-Consistent High Fidelity Bipolar EEG Generation". "Physical validity" no longer appears anywhere; we call the property montage consistency.
  • Recording vs. view: the introduction now states that EEG is recorded against a common reference and read in a bipolar montage derived from it, and that the bipolar channels are constrained by how they are derived, not by how the signal is recorded.
  • Scope: the abstract and conclusion now refer only to generators of bipolar EEG.
  • Framing: we removed "categorical failure", "has gone unnoticed", "EEG is not an arbitrary multichannel signal", and the description of the identities as physical laws.

Concerns addressed below:

  1. The identities are arithmetic consequences of the derivation, not EEG physics. Agreed, fully. They hold for any signal derived this way.

  2. They only concern generators that model the bipolar view. Agreed, and the paper now says so.

  3. Generating common-reference EEG and then deriving the bipolar view is also valid. Agreed, so validity alone does not set MontageFlow apart. What does is fidelity at equal validity, and the experiment in our previous reply shows it: both routes use our latent-conditioned model and are exactly consistent. Only the way consistency is enforced differs. Test TS-FID is shown below (with % reduction) :

RouteTUSZTUEVTUAB
MontageFlow66.0115.754.7
Trained on common-reference EEG (bipolar channels derived afterwards)102.9 (−36%)125.2 (−8%)118.4 (−54%)

We will add this comparison to the paper.

  1. Per-hemisphere vs. whole-head average. You are right. The 16-channel montage does not connect the two hemispheres, so a whole-head average cannot be derived from it (for real windows either).

Generating the whole-head average directly. The common-average experiment in our previous reply already used the whole-head average of all 16 electrodes. Test TS-FID is shown below (with % reduction) :

Model trained on the whole-head averageTUSZTUEVTUAB
Without our projector101.7107.4100.7
With our (zero-sum) projector66.0 (−35%)100.4 (−7%)75.7 (−25%)

So when the whole-head average is the target, the right route is to generate it directly with our projector.


Thank you again. Your review made the paper more precise, and we are grateful for it. If these changes address your concerns, we would be grateful if you would consider updating your score. We are happy to adjust the framing further if anything still reads as overclaiming.

Thank you, and we are glad your concerns are resolved. We have made both changes in the revised PDF:

The abstract now reads: “Across 3 large-scale benchmarks in the bipolar montage, our model improves TS-FID by 1.6–3.0× over the closest rival while ensuring montage consistency.” Appendix D now reports these comparisons, with the real-vs-real reference (Tables 7 and 8).

Thank you again for helping us make the paper more precise.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.