pith. sign in

arxiv: 2605.21280 · v1 · pith:GYLDR2MZnew · submitted 2026-05-20 · 💻 cs.CV

Let EEG Models Learn EEG

Yifan Wang , Yijia Ma , Wen Li , Chenyu You This is my paper

Pith reviewed 2026-05-21 05:08 UTC · model grok-4.3

classification 💻 cs.CV
keywords EEG generationflow matchingcontinuous trajectoriesneural dynamicsgenerative modelsspectral constraintstemporal stationarity
0
0 comments X

The pith

EEG generation improves when models learn continuous trajectories of raw signals instead of discrete denoising steps.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that most EEG generators rely on discrete denoising, which breaks the continuous temporal dynamics and spectral properties of real brain signals. JET instead uses conditional flow matching to learn a smooth vector field that moves noise along continuous paths directly to EEG data distributions. Constraints are added to keep the learned dynamics aligned with spectral structure, temporal stationarity, and basic signal statistics. If the approach holds, it would let researchers create large volumes of realistic synthetic EEG for model training while respecting privacy and data limits.

Core claim

The paper claims that modeling EEG as raw sequences evolving along continuous trajectories via conditional flow matching, combined with constraints that preserve spectral structure, temporal stationarity, and signal-level statistics, produces generated signals whose long-range dependencies and transient dynamics match real neural activity more closely than discrete denoising methods do.

What carries the argument

Conditional flow matching that learns a smooth vector field transporting noise to the EEG data distribution, subject to constraints on spectral structure, temporal stationarity, and signal statistics.

If this is right

  • JET reduces TS-FID by over 40 percent versus strong baselines on three large-scale EEG benchmarks.
  • Generated signals maintain long-range temporal dependencies that discrete methods typically lose.
  • The model captures key structural properties of neural dynamics without domain-specific representations.
  • The framework scales to high-volume synthetic data production for downstream neural modeling tasks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same continuous-trajectory approach could be tested on other continuous biosignals such as ECG or MEG.
  • Synthetic EEG produced this way might serve as augmentation data for training brain-computer interface models.
  • The results suggest that any time-series generation task with inherent continuity may benefit from replacing discrete diffusion steps with flow matching plus structure-preserving constraints.

Load-bearing premise

The introduced constraints will keep the learned continuous vector field consistent with real EEG properties without introducing artifacts or losing long-range dependencies.

What would settle it

Generate long EEG traces with JET and compare their power spectral density and autocorrelation functions against real recordings; systematic mismatches in low-frequency content or stationarity statistics would show the constraints failed to enforce consistency.

Figures

Figures reproduced from arXiv: 2605.21280 by Chenyu You, Wen Li, Yifan Wang, Yijia Ma.

Figure 1
Figure 1. Figure 1: Toy comparison of generative modeling paradigms. We visualize the learned distributions on three different manifolds. GANs map latent noise directly to data space, diffusion models rely on stochastic denoising dynamics, while flow matching learns a smooth time-dependent vector field whose integral curves realize optimal transport from a dispersed source to the target distribution. eration. JET models multi… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of JET pipeline. The model generates multi-channel EEG by learning a continuous vector field v(xt, t) via flow matching, conditioned on pathological states and regularized by structure-preserving constraints that encode spectral, temporal, and statistical properties of EEG signals. constraints that regularize the generative flow to preserve spectral, temporal, and statistical structure of EEG … view at source ↗
Figure 4
Figure 4. Figure 4: Evaluation of temporal dynamics. Temporal evolution of signal amplitude distributions. JET captures non-stationary energy fluctuations over time while maintaining stable amplitude statistics, avoiding temporal drift and variance collapse. Method W1(slope) ↓ W1(Dµ) ↓ W1(Dσ) ↓ Real split 0.008 0.012 0.010 EEG-GAN 0.065 0.078 0.071 Vanilla Diff. 0.051 0.063 0.058 JET 0.015 0.021 0.018 [PITH_FULL_IMAGE:figure… view at source ↗
Figure 5
Figure 5. Figure 5: Analysis of amplitude distributions. Log-density his￾tograms confirm that JET accurately reconstructs the non-Gaussian, heavy-tailed distributions of raw EEG, effectively covering the di￾verse amplitude ranges of pathological recordings. 0 5 10 15 20 25 30 35 40 TUAB 0 5 10 15 20 25 30 35 40 TUEV 0 5 10 15 20 25 30 35 40 Frequency (Hz) TUSZ Real median Generated median [PITH_FULL_IMAGE:figures/full_fig_p0… view at source ↗
Figure 6
Figure 6. Figure 6: Analysis of population diversity. The extensive over￾lap in spectral variance envelopes confirms that JET captures the complex variances of the population, avoiding the mode collapse typical of Gaussian-based priors. µ last c | and Dσ = 1 C P c |σ first c − σ last c | between the first and last quarter of each segment. We then compare the distri￾bution of each statistic against held-out real data via the 1… view at source ↗
Figure 7
Figure 7. Figure 7 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Gaussian noise: amplitude distribution. The generated signals closely match the ground truth log-density histograms, effectively capturing the heavy-tailed distribution of EEG amplitudes across all datasets. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Gaussian noise: power spectral density. The model accurately reconstructs the 1/f spectral slope and distinct oscillatory peaks, demonstrating high spectral fidelity. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Gaussian noise: spectral distribution stability. The shaded percentile bands of the generated spectra overlap tightly with the real data, indicating that JET captures the full diversity of population-level variance. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Gaussian noise: temporal stationarity. Z-normalized temporal envelopes exhibit stable, non-stationary variance dynamics that mirror the natural evolution of biological signals without drift. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Zero noise: amplitude distribution. Initializing with zero noise leads to severe mode collapse, visualized as sharp, unnatural spikes in the histograms and a failure to cover the marginal distribution. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Zero noise: power spectral density. The spectra exhibit significant distortion, with flattened slopes and mismatched energy levels, confirming the inability to model scale-free dynamics from a singularity. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Zero noise: spectral distribution stability. The generated variance bands are either significantly collapsed (too narrow) or distorted, failing to represent the natural spectral diversity of the dataset. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Zero Noise: temporal stationarity. The temporal traces display erratic fluctuations or flat-lining, reflecting the failure of the flow trajectory to evolve into valid temporal patterns. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_15.png] view at source ↗
read the original abstract

High-fidelity EEG generation is critical for alleviating data scarcity and addressing privacy constraints in large-scale neural modeling. Despite recent progress, most existing approaches formulate EEG generation via discrete denoising objectives, which inadequately reflect the inherently continuous temporal dynamics and spectral structure of neural activity. As a result, these methods often struggle to preserve long-range temporal dependencies and exhibit mismatches in the spectral and temporal structure of the generated signals. In this work, we argue that effective EEG generation requires models that operate directly on the continuous evolution of neural signals. We introduce Just EEG Transformer (JET), a generative framework based on conditional flow matching that models EEG as raw sequences evolving along continuous trajectories. By learning a smooth vector field that transports noise to the EEG data distribution, JET captures temporal continuity and transient dynamics without relying on discretized denoising schemes or domain-specific representations. To ensure that the learned dynamics remain consistent with key properties of EEG signals, we introduce principled constraints that preserve spectral structure, temporal stationarity, and signal-level statistics. Across three large-scale benchmarks, JET consistently achieves state-of-the-art performance, reducing TS-FID by over 40% compared to strong baselines. Extensive analyses show that JET captures key structural properties of neural dynamics, providing a scalable and principled approach to EEG generation. Project page: https://y-research-sbu.github.io/JET/ .

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper introduces Just EEG Transformer (JET), a generative framework for high-fidelity EEG synthesis based on conditional flow matching. It models EEG signals as raw sequences evolving along continuous trajectories by learning a smooth vector field, augmented with three principled constraints to preserve spectral structure, temporal stationarity, and signal-level statistics. The central claim is that this approach outperforms discrete denoising baselines, achieving state-of-the-art results with over 40% reduction in TS-FID across three large-scale benchmarks while better capturing long-range temporal dependencies and neural dynamics.

Significance. If the empirical claims hold after verification, the work offers a principled shift from discrete to continuous modeling for EEG generation. This could meaningfully address data scarcity and privacy issues in neural signal processing by producing trajectories that respect EEG-specific properties, with potential downstream benefits for augmentation in BCI and clinical modeling tasks. The emphasis on flow matching and explicit constraints is a strength if supported by ablations and guarantees.

major comments (3)
  1. [Abstract and §3] Abstract and §3: The central performance claim of >40% TS-FID reduction is presented without accompanying statistical tests, confidence intervals, or error analysis. This makes it impossible to determine whether the improvement is robust or attributable to the proposed constraints versus the base conditional flow matching architecture.
  2. [§3.2] §3.2 (Constraints): The spectral, stationarity, and signal-level constraints are introduced as auxiliary losses or regularizers, but no derivation, weighting schedule, or bound is provided to guarantee that the resulting ODE trajectories remain on the manifold of valid EEG signals. This directly bears on the weakest assumption that long-range dependencies and transient events are preserved without artifacts.
  3. [§4] §4 (Experiments): No ablation studies isolating the contribution of each constraint, no implementation details (e.g., vector field parameterization, ODE solver tolerances), and no comparison against ablated versions of the flow-matching objective alone are reported. These omissions prevent verification that the reported gains stem from the full JET framework.
minor comments (2)
  1. [§2] Notation for the vector field and conditioning variables should be defined more explicitly in §2 to avoid ambiguity when reading the constraint formulations.
  2. [Figures] Figure captions and axis labels in the qualitative results could be expanded to indicate which constraint is active in each panel.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key opportunities to strengthen the empirical validation and methodological details of the JET framework. We respond to each major comment below and describe the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3: The central performance claim of >40% TS-FID reduction is presented without accompanying statistical tests, confidence intervals, or error analysis. This makes it impossible to determine whether the improvement is robust or attributable to the proposed constraints versus the base conditional flow matching architecture.

    Authors: We agree that statistical analysis is necessary to support the performance claims. In the revised manuscript we will report TS-FID means accompanied by standard deviations computed over multiple independent runs with different random seeds. We will also add statistical significance tests (paired t-tests and Wilcoxon signed-rank tests) between JET and the baselines on each benchmark to demonstrate that the observed reductions are robust and not explained by variance alone. revision: yes

  2. Referee: [§3.2] §3.2 (Constraints): The spectral, stationarity, and signal-level constraints are introduced as auxiliary losses or regularizers, but no derivation, weighting schedule, or bound is provided to guarantee that the resulting ODE trajectories remain on the manifold of valid EEG signals. This directly bears on the weakest assumption that long-range dependencies and transient events are preserved without artifacts.

    Authors: The constraints are introduced as soft penalty terms within the conditional flow-matching objective to encourage preservation of EEG-specific properties. Deriving a rigorous bound that keeps every ODE trajectory exactly on the valid EEG manifold is a non-trivial theoretical question that goes beyond the scope of the current empirical study. In the revision we will expand §3.2 with the concrete weighting schedule (including the values of the regularization coefficients and any annealing schedule used during training) and provide an empirical analysis of how the constraints affect the learned vector field and the generated trajectories. We will also discuss observed artifacts and the practical mitigation achieved by the combined objective. revision: partial

  3. Referee: [§4] §4 (Experiments): No ablation studies isolating the contribution of each constraint, no implementation details (e.g., vector field parameterization, ODE solver tolerances), and no comparison against ablated versions of the flow-matching objective alone are reported. These omissions prevent verification that the reported gains stem from the full JET framework.

    Authors: We will add a dedicated ablation subsection and an expanded appendix. The new experiments will include: (i) JET with each constraint removed individually, (ii) a pure conditional flow-matching baseline without any of the three constraints, and (iii) comparisons against discrete denoising models augmented with the same constraints. We will also document the vector-field network architecture, the ODE solver (including tolerances such as atol and rtol), and all training hyperparameters. These additions will make the contribution of each component verifiable. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on established conditional flow matching plus independently motivated constraints

full rationale

The paper's core framework is conditional flow matching applied to continuous EEG trajectories, an approach drawn from prior literature rather than defined in terms of the target EEG properties. The added constraints on spectral structure, temporal stationarity, and signal statistics are presented as auxiliary regularizers motivated by known EEG characteristics; they are not shown to be fitted to the model's own outputs or to reduce the performance metric by construction. No self-citation chain is invoked to justify uniqueness or to smuggle in an ansatz, and the reported TS-FID improvements are framed as empirical outcomes rather than tautological predictions. The derivation chain therefore remains self-contained against external benchmarks and does not collapse to its inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Relies on standard flow-matching assumptions plus newly introduced constraints whose precise mathematical form and enforcement are not detailed in the abstract.

axioms (1)
  • domain assumption Conditional flow matching can faithfully transport noise to the distribution of raw EEG sequences while preserving long-range temporal dependencies.
    Central to the claim that continuous trajectories outperform discrete denoising for EEG.

pith-pipeline@v0.9.0 · 5763 in / 1270 out tokens · 48853 ms · 2026-05-21T05:08:24.165352+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

117 extracted references · 117 canonical work pages · 10 internal anchors

  1. [1]

    Langley , title =

    P. Langley , title =. Proceedings of the 17th International Conference on Machine Learning (ICML 2000) , address =. 2000 , pages =

    work page 2000

  2. [2]

    T. M. Mitchell. The Need for Biases in Learning Generalizations. 1980

    work page 1980

  3. [3]

    M. J. Kearns , title =

    work page

  4. [4]

    Machine Learning: An Artificial Intelligence Approach, Vol. I. 1983

    work page 1983

  5. [5]

    R. O. Duda and P. E. Hart and D. G. Stork. Pattern Classification. 2000

    work page 2000

  6. [6]

    Suppressed for Anonymity , author=

    work page

  7. [7]

    Newell and P

    A. Newell and P. S. Rosenbloom. Mechanisms of Skill Acquisition and the Law of Practice. Cognitive Skills and Their Acquisition. 1981

    work page 1981

  8. [8]

    A. L. Samuel. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development. 1959

    work page 1959

  9. [9]

    arXiv preprint arXiv:2405.18765 , year=

    Large brain model for learning generic representations with tremendous EEG data in BCI , author=. arXiv preprint arXiv:2405.18765 , year=

    work page arXiv

  10. [10]

    arXiv preprint arXiv:2401.10278 , year=

    Eegformer: Towards transferable and interpretable large-scale eeg foundation model , author=. arXiv preprint arXiv:2401.10278 , year=

    work page arXiv

  11. [11]

    2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=

    Neuro-gpt: Towards a foundation model for eeg , author=. 2024 IEEE International Symposium on Biomedical Imaging (ISBI) , pages=. 2024 , organization=

    work page 2024

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Eegpt: Pretrained transformer for universal and reliable representation of eeg signals , author=. Advances in Neural Information Processing Systems , volume=

    work page

  13. [13]

    arXiv preprint arXiv:2305.10351 , year=

    BIOT: Cross-data biosignal learning in the wild , author=. arXiv preprint arXiv:2305.10351 , year=

    work page arXiv

  14. [14]

    Dr eamDif- fusion: Generating high-quality images from brain EEG sign als,

    Dreamdiffusion: Generating high-quality images from brain eeg signals , author=. arXiv preprint arXiv:2306.16934 , year=

    work page arXiv

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Seeing beyond the brain: Conditional diffusion model with sparse masked modeling for vision decoding , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  16. [16]

    arXiv preprint arXiv:2409.00101 , year=

    NeuroLM: A universal multi-task foundation model for bridging the gap between language and EEG signals , author=. arXiv preprint arXiv:2409.00101 , year=

    work page arXiv

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Brant: Foundation model for intracranial neural signal , author=. Advances in Neural Information Processing Systems , volume=

    work page

  18. [18]

    arXiv preprint arXiv:2510.12515 , year=

    HEAR: An EEG Foundation Model with Heterogeneous Electrode Adaptive Representation , author=. arXiv preprint arXiv:2510.12515 , year=

    work page arXiv

  19. [19]

    Zhou, Y ., Wu, J., Ren, Z., Yao, Z., Lu, W., Peng, K., Zheng, Q., Song, C., Ouyang, W., and Gou, C

    CSBrain: A Cross-scale Spatiotemporal Brain Foundation Model for EEG Decoding , author=. arXiv preprint arXiv:2506.23075 , year=

    work page arXiv

  20. [20]

    Cbramod: A criss-cross brain foundation model for eeg decoding.arXiv preprint arXiv:2412.07236, 2024a

    Cbramod: A criss-cross brain foundation model for eeg decoding , author=. arXiv preprint arXiv:2412.07236 , year=

    work page arXiv

  21. [21]

    arXiv preprint arXiv:2511.14196 , year=

    MindCross: Fast New Subject Adaptation with Limited Data for Cross-subject Video Reconstruction from Brain Signals , author=. arXiv preprint arXiv:2511.14196 , year=

    work page arXiv

  22. [22]

    arXiv preprint arXiv:2504.11936 , year=

    Mind2matter: Creating 3d models from eeg signals , author=. arXiv preprint arXiv:2504.11936 , year=

    work page arXiv

  23. [23]

    arXiv preprint arXiv:2204.12440 , year=

    Neuro-bert: Rethinking masked autoencoding for self-supervised neurological pretraining , author=. arXiv preprint arXiv:2204.12440 , year=

    work page arXiv

  24. [24]

    U., Kreiman, G., Katz, B., Cases, I., and Barbu, A

    BrainBERT: Self-supervised representation learning for intracranial recordings , author=. arXiv preprint arXiv:2302.14367 , year=

    work page arXiv

  25. [25]

    arXiv preprint arXiv:2308.13234 , year=

    Decoding natural images from eeg for object recognition , author=. arXiv preprint arXiv:2308.13234 , year=

    work page arXiv

  26. [26]

    IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=

    EEG conformer: Convolutional transformer for EEG decoding and visualization , author=. IEEE Transactions on Neural Systems and Rehabilitation Engineering , volume=. 2022 , publisher=

    work page 2022

  27. [27]

    2018 , month =

    Lawhern, Vernon J and Solon, Amelia J and Waytowich, Nicholas R and Gordon, Stephen M and Hung, Chou P and Lance, Brent J , title =. 2018 , month =. doi:10.1088/1741-2552/aace8c , url =

    work page doi:10.1088/1741-2552/aace8c 2018

  28. [28]

    Human brain mapping , volume=

    Deep learning with convolutional neural networks for EEG decoding and visualization , author=. Human brain mapping , volume=. 2017 , publisher=

    work page 2017

  29. [29]

    EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals

    EEG-GAN: Generative adversarial networks for electroencephalograhic (EEG) brain signals , author=. arXiv preprint arXiv:1806.01875 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    Proceedings of the 25th ACM international conference on Multimedia , pages=

    Brain2image: Converting brain signals into images , author=. Proceedings of the 25th ACM international conference on Multimedia , pages=

    work page

  31. [31]

    arXiv preprint arXiv:2309.14030 , year=

    Dewave: Discrete eeg waves encoding for brain dynamics to text translation , author=. arXiv preprint arXiv:2309.14030 , year=

    work page arXiv

  32. [32]

    Proceedings of the AAAI conference on artificial intelligence , year=

    Towards voice reconstruction from EEG during imagined speech , author=. Proceedings of the AAAI conference on artificial intelligence , year=

    work page

  33. [33]

    arXiv preprint arXiv:2308.02510 , year=

    Seeing through the brain: image reconstruction of visual perception from human brain signals , author=. arXiv preprint arXiv:2308.02510 , year=

    work page arXiv

  34. [34]

    Frontiers in Human Neuroscience , volume=

    BENDR: Using transformers and a contrastive self-supervised learning task to learn from massive amounts of EEG data , author=. Frontiers in Human Neuroscience , volume=. 2021 , publisher=

    work page 2021

  35. [35]

    IEEE Transactions on Cognitive and Developmental Systems , year=

    Data augmentation for seizure prediction with generative diffusion model , author=. IEEE Transactions on Cognitive and Developmental Systems , year=

    work page

  36. [36]

    Visual neural decod- ing via improved visual-eeg semantic consistency

    Visual neural decoding via improved visual-EEG semantic consistency , author=. arXiv preprint arXiv:2408.06788 , year=

    work page arXiv

  37. [37]

    Back to Basics: Let Denoising Generative Models Denoise

    Back to basics: Let denoising generative models denoise , author=. arXiv preprint arXiv:2511.13720 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Scalable diffusion models with transformers , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

    work page

  39. [39]

    Flow Matching for Generative Modeling

    Flow matching for generative modeling , author=. arXiv preprint arXiv:2210.02747 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  40. [40]

    2025 , eprint=

    Mean Flows for One-step Generative Modeling , author=. 2025 , eprint=

    work page 2025

  41. [41]

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow

    Flow straight and fast: Learning to generate and transfer data with rectified flow , author=. arXiv preprint arXiv:2209.03003 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  42. [42]

    2024 , eprint=

    SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers , author=. 2024 , eprint=

    work page 2024

  43. [43]

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions

    Stochastic interpolants: A unifying framework for flows and diffusions , author=. arXiv preprint arXiv:2303.08797 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

    work page

  45. [45]

    International Conference on Learning Representations , year=

    Denoising Diffusion Implicit Models , author=. International Conference on Learning Representations , year=

    work page

  46. [46]

    2021 , eprint=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. 2021 , eprint=

    work page 2021

  47. [47]

    2021 , eprint=

    High-Resolution Image Synthesis with Latent Diffusion Models , author=. 2021 , eprint=

    work page 2021

  48. [48]

    IEEE transactions on medical imaging , year=

    CT super-resolution GAN constrained by the identical, residual, and cycle learning ensemble (GAN-CIRCLE) , author=. IEEE transactions on medical imaging , year=

    work page

  49. [49]

    arXiv preprint arXiv:2511.19917 , year=

    Scale Where It Matters: Training-Free Localized Scaling for Diffusion Models , author=. arXiv preprint arXiv:2511.19917 , year=

    work page arXiv

  50. [50]

    Physics in Medicine & Biology , year=

    MARGANVAC: metal artifact reduction method based on generative adversarial network with variable constraints , author=. Physics in Medicine & Biology , year=

    work page

  51. [51]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    Coma: Compositional human motion generation with multi-modal agents , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

    work page

  52. [52]

    Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

    Hybrid neural diffeomorphic flow for shape representation and generation via triplane , author=. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , year=

    work page

  53. [53]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

    Medgen3d: A deep generative framework for paired 3d image and mask generation , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

    work page

  54. [54]

    International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

    Pre-trained diffusion models for plug-and-play medical image enhancement , author=. International Conference on Medical Image Computing and Computer-Assisted Intervention , year=

    work page

  55. [55]

    2023 , eprint=

    Consistency Models , author=. 2023 , eprint=

    work page 2023

  56. [56]

    2023 , eprint=

    Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference , author=. 2023 , eprint=

    work page 2023

  57. [57]

    International Conference on Learning Representations , year=

    InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation , author=. International Conference on Learning Representations , year=

    work page

  58. [58]

    The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    MaskGIT: Masked Generative Image Transformer , author=. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    work page

  59. [59]

    2025 , eprint=

    DC-AR: Efficient Masked Autoregressive Image Generation with Deep Compression Hybrid Tokenizer , author=. 2025 , eprint=

    work page 2025

  60. [60]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    Fisher Flow Matching for Generative Modeling over Discrete Data , author=. The Thirty-eighth Annual Conference on Neural Information Processing Systems , year=

    work page

  61. [61]

    2014 , eprint=

    Generative Adversarial Networks , author=. 2014 , eprint=

    work page 2014

  62. [62]

    2017 , eprint=

    Improved Training of Wasserstein GANs , author=. 2017 , eprint=

    work page 2017

  63. [63]

    2019 , eprint=

    A Style-Based Generator Architecture for Generative Adversarial Networks , author=. 2019 , eprint=

    work page 2019

  64. [64]

    Large Scale

    Andrew Brock and Jeff Donahue and Karen Simonyan , booktitle=. Large Scale. 2019 , url=

    work page 2019

  65. [65]

    NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    Classifier-Free Diffusion Guidance , author=. NeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications , year=

    work page 2021

  66. [66]

    Tero Karras and Miika Aittala and Timo Aila and Samuli Laine , title =. Proc. NeurIPS , year =

    work page

  67. [67]

    2022 , eprint=

    Hierarchical Text-Conditional Image Generation with CLIP Latents , author=. 2022 , eprint=

    work page 2022

  68. [68]

    2022 , eprint=

    Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding , author=. 2022 , eprint=

    work page 2022

  69. [69]

    Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , month =

    Shi, Yingdong and Li, Changming and Wang, Yifan and Zhao, Yongxiang and Pang, Anqi and Yang, Sibei and Yu, Jingyi and Ren, Kan , title =. Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR) , month =. 2025 , pages =

    work page 2025

  70. [70]

    Ouroboros: Single-step Diffusion Models for Cycle-consistent Forward and Inverse Rendering

    Ouroboros: Single-step Diffusion Models for Cycle-consistent Forward and Inverse Rendering , author=. arXiv preprint arXiv:2508.14461 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  71. [71]

    IEEE International Conference on Computer Vision (ICCV) , year=

    Adding Conditional Control to Text-to-Image Diffusion Models , author=. IEEE International Conference on Computer Vision (ICCV) , year=

    work page

  72. [72]

    Advances in Neural Information Processing Systems , editor=

    On Density Estimation with Diffusion Models , author=. Advances in Neural Information Processing Systems , editor=. 2021 , url=

    work page 2021

  73. [73]

    2021 , url=

    Improved Denoising Diffusion Probabilistic Models , author=. 2021 , url=

    work page 2021

  74. [74]

    Thirty-seventh Conference on Neural Information Processing Systems , year=

    Cold Diffusion: Inverting Arbitrary Image Transforms Without Noise , author=. Thirty-seventh Conference on Neural Information Processing Systems , year=

    work page

  75. [75]

    Transactions on Machine Learning Research , issn=

    Soft Diffusion: Score Matching with General Corruptions , author=. Transactions on Machine Learning Research , issn=. 2023 , url=

    work page 2023

  76. [76]

    2018 , eprint=

    Neural Discrete Representation Learning , author=. 2018 , eprint=

    work page 2018

  77. [77]

    arXiv preprint arXiv:2602.21333 , year=

    HorizonForge: Driving Scene Editing with Any Trajectories and Any Vehicles , author=. arXiv preprint arXiv:2602.21333 , year=

    work page arXiv

  78. [78]

    Diffusion Transformers with Representation Autoencoders

    Diffusion transformers with representation autoencoders , author=. arXiv preprint arXiv:2510.11690 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  79. [79]

    Frontiers in Neuroscience , VOLUME=

    Obeid, Iyad and Picone, Joseph , TITLE=. Frontiers in Neuroscience , VOLUME=. 2016 , URL=. doi:10.3389/fnins.2016.00196 , ISSN=

    work page doi:10.3389/fnins.2016.00196 2016

  80. [80]

    Journal of Machine Learning Research , volume=

    Normalizing flows for probabilistic modeling and inference , author=. Journal of Machine Learning Research , volume=

    work page

Showing first 80 references.