Pith. sign in

REVIEW 4 major objections 5 minor 84 references

One-step generative surrogates can match neural-operator speed while preserving long-horizon turbulence statistics.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 14:43 UTC pith:IMLN6CJR

load-bearing objection A genuinely new one-step autoregressive generative surrogate with solid ablations, but the missing rolling generative baselines and a stretched 'provably cannot' claim need referee attention. the 4 major comments →

arxiv 2605.05540 v2 pith:IMLN6CJR submitted 2026-05-07 cs.LG physics.flu-dyn

Autoregressive One-Step Generative Modeling for Dynamical System Forecasting

Tianyue Yang , Xiao Xue This is my paper
classification cs.LG physics.flu-dyn
keywords one-step generative modelingdynamical system forecastingMeanFlowturbulence surrogatesautoregressive generationlong-horizon statisticstemporal increment consistencyneural operators
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces MeLISA, a stochastic surrogate for forecasting physical dynamics that produces each multi-frame forecast block in a single network evaluation, without latent encoders or iterative diffusion solvers. Its aim is to show that a one-step generative transition kernel can be as fast as deterministic neural operators while better preserving long-horizon statistical structure—energy spectra, turbulent kinetic energy, and mixing rate—on 256×256 Kolmogorov flow and 192×192 turbulent channel flow. The method couples two training objectives: Window-Consistency MeanFlow, which learns conditional generation from partially masked temporal windows, and Time Increment Consistency, which penalizes mismatch in multi-lag temporal increments. A sympathetic reader would take the paper to establish that one-step stochastic forecasting is no longer incompatible with physical realism in turbulent regimes.

Core claim

MeLISA's central claim is that a latent-free, one-step generative model can serve as an autoregressive surrogate for turbulent physical dynamics: each forecast block is produced by a single denoising evaluation of a pixel-space MeanFlow network, conditioned on a partially masked temporal window, and a Time Increment Consistency loss forces the generated blocks to match ground-truth multi-lag increments. On 256×256 Kolmogorov flow and 192×192 turbulent channel-flow slices, MeLISA reportedly matches or exceeds neural-operator baselines on short-term relative L2 error and SSIM while achieving better power-spectral-density discrepancy, turbulent kinetic energy difference, and mixing-rate differe

What carries the argument

The engine is the blockwise stochastic transition kernel built from Window-Consistency MeanFlow (WinC-MF). WinC-MF extends pixel MeanFlow from single frames to temporal windows: a denoiser receives a noisy window plus a mask revealing only some frames, and must reconstruct the whole window, so uncertainty over masked frames survives and the model cannot collapse to copying. The Time Increment Consistency (TIC) regularizer then compares predicted against true increments x_{τ+w}−x_τ for lags w=1..W−1; at second order these increments are equivalent to temporal covariance decay, so TIC supplies the long-horizon constraint that pointwise losses cannot. Inference applies the denoiser once per blo

Load-bearing premise

The load-bearing premise is that matching the first five lagged increments of a short training window forces the full multi-hundred-frame rollout to reproduce the true energy spectra, kinetic-energy map, and mixing rate—a transfer the paper proves only at second order for stationary processes.

What would settle it

Train MeLISA on a dataset whose dynamics carry significant memory beyond the five training lags, then compare full-trajectory autocorrelation at lags 6 through the rollout horizon against ground truth; if autocorrelation diverges while the TIC loss at lags 1–5 is minimized, the short-lag-to-long-horizon bridge fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is right, generative surrogates no longer need multi-step solvers to be statistically reliable; one evaluation per block is enough.
  • Autoregressive neural-operator surrogates can be replaced by one-step stochastic kernels with comparable wall-clock speed and better long-horizon spectral and kinetic-energy recovery.
  • TIC acts as a cheap finite-lag regularizer that constrains temporal covariance without explicitly computing a spectral loss during training.
  • Larger MeLISA variants show improving long-term statistical metrics even as short-term pointwise metrics saturate, suggesting scaling shifts behavior toward context-aware statistical fidelity.
  • Rollouts of thousands of frames remain stable for MeLISA variants, whereas some neural-operator baselines fail to roll out to the full in-distribution trajectory length.

Where Pith is reading between the lines

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

  • TIC's success suggests that other stochastic or deterministic surrogates could adopt finite-lag increment matching as a lightweight substitute for explicit spectral penalties; this is an editorial extrapolation, not a paper claim.
  • Because inference decouples forecast-block size from the trained window length, MeLISA could in principle be deployed at multiple temporal scales by resizing the rollout block, a flexibility the paper leaves implicit.
  • The masked-window objective resembles self-supervised prediction, which hints that further scaling and larger training corpora may produce stronger context utilization in generative surrogate models; the paper frames this as a possibility, not an established result.
  • If short-lag increment matching genuinely controls long-horizon covariance decay, TIC may transfer to other partially observed or non-Markovian dynamical systems beyond turbulence, but this transfer is not demonstrated here.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes MeLISA, a latent-free autoregressive generative surrogate for high-dimensional physical dynamics. It extends pixel-space MeanFlow to a window-conditioned, masked spatiotemporal setting (WinC-MF) and adds a Time Increment Consistency (TIC) regularizer over finite-lag differences (Eq. 20). Inference is one function evaluation per forecast block. The method is evaluated on 192x192 turbulent channel flow and 256x256 Kolmogorov flow, against FNO, UNO, and Local-FNO, with short-term RL2/SSIM and long-term PSDD/TKED/MRD metrics. The central claim is that MeLISA is the first high-resolution one-step generative surrogate whose long-horizon statistical fidelity is comparable to deterministic neural operators while preserving fast rollout.

Significance. If the claims hold, the paper makes a useful contribution: a 1-NFE stochastic autoregressive model that avoids latent encoders and iterative denoising, with evidence of better spectral and temporal statistics than neural-operator baselines. The manuscript is generally careful about training/inference costs and includes ablation studies (D.1, D.2), a long-rollout stress test (D.3), and probabilistic CRPS evaluation (D.4). The main methodological novelty, WinC-MF with TIC, is clearly motivated and the theoretical appendix connects TIC to covariance decay under stationarity. However, the evaluation is narrower than the claim: the closest generative baselines are discussed but not benchmarked, the TIC short-lag/long-lag transfer is not established, and there are internal numerical inconsistencies in the key ablation table. The central idea is defensible but the current evidence base is not yet sufficient to support the abstract-level claim.

major comments (4)
  1. [Tables 1–2 and Appendix B] MeLISA is compared only to deterministic neural operators (FNO, UNO, Local-FNO). The closest family — rolling diffusion/flow-matching surrogates such as ERDM, RDM, DYffusion, GenCFD, and rectified-flow baselines — is reviewed in Appendix B and listed in Table 8, but none are run on these benchmarks. Since the paper's central contribution is one-step generative forecasting and the abstract claims 'performance comparable to state-of-the-art deterministic surrogates,' the 1-NFE advantage is not enough: the reader cannot tell whether MeLISA trades away calibration or long-horizon diversity relative to multi-step generative methods. Please add at least one or two rolling generative baselines (or publish existing numbers on the same datasets) and report CRPS for them as well.
  2. [Eq. (20), Table 4, App. C.2, App. F.3] TIC is trained only on lags w=1,...,5 with weights (0.4, 0.5, 0.8, 1.1, 1.2), while MRD fits the autocorrelation decay over the first 20 frames (App. F.3, Eq. (80)). Appendix C.2, Eq. (47), shows that squared finite-lag increments equal 2tr(Γ(0)−Γ(w)) under second-order stationarity, but it does not prove that controlling lags 1–5 transfers to lags 6–20, to the full spectrum, or to the invariant measure. Because the diagnostic and the loss share the same short-lag range, the MRD improvements in Tables 1–2 may partly reflect direct optimization of the evaluation quantity. Please add held-out diagnostics: for example, report autocorrelation error on lags 6–20 separately, and include a KF256 variant trained with TIC only on lags 6–20 (or with TIC weights that exclude the diagnostic lags). Without such a test, the long-horizon-transfer claim overreaches.
  3. [Appendix D.2, Table 9 vs Table 2] Table 9 reports the full MeLISA-∆-M on KF256 with PSDD=0.540, TKED=1.73, MRD=0.211, whereas Table 2 reports the same model with PSDD=0.445(78), TKED=4.24(64), MRD=0.086(20). These differ by factors of roughly 1.2–2.5, well outside the quoted uncertainties. If Table 9 is the same trained model, the discrepancy needs an explanation (checkpoint, evaluation subset, or seed); if it is a new run, the 'same optimization setting' must be stated and uncertainties given. This ablation is the primary evidence that TIC, rather than other properties of the generative model, drives long-horizon improvements, so the inconsistency substantially weakens that evidence.
  4. [Sec. 1 contribution bullet, App. C.2 Eqs. (56)–(57)] The contribution bullet states that TIC supplies 'long-horizon constraints that pointwise state-reconstruction losses provably cannot.' Appendix C.2 shows that the minimizer of a frame-wise squared loss is the conditional mean, and that such a loss does not directly constrain finite-lag increments. However, 'provably cannot' is stronger than what is shown: a multi-frame state-reconstruction loss on windows, or a loss with explicit spectral weighting, could constrain the same second-order structure. Please weaken the claim, or state and prove a precise impossibility result for a defined class of pointwise losses.
minor comments (5)
  1. [Sec. 5.2, Table 2] The scaling narrative is mixed: MRD increases from 0.086 (∆-M) to 0.130 (∆-B), and RL2 is non-monotonic. The text acknowledges non-monotonicity but the conclusion that the DiT path 'provides a scalable path' should be qualified by noting that not all long-term statistics improve with scale.
  2. [Appendix D.3, Fig. 10] The 9998-frame stress test is reported only qualitatively with white frames indicating NaN or unavailable values. A quantitative metric at, say, 1000 or 9998 frames (PSDD, TKED, or RL2) would make the stability claim more convincing.
  3. [Appendix B, final paragraph] The statement that MeLISA has 'no strict counterpart' is too strong. Table 8 lists one-step latent-free methods (e.g., DYffusion and rectified-flow baselines) that are close in spirit. The novelty should be framed as the specific combination of window-consistency masking and finite-lag regularization, not as the absence of any related work.
  4. [Table 4] The TIC lag weights and the masking rate (0.8) are tuned hyperparameters. No sensitivity analysis is reported for these, although the method's behavior depends on them. A small sensitivity study (e.g., varying the masking rate in [0.6, 0.9] or shifting TIC weights) would strengthen the claims of robustness.
  5. [App. F.2, Eq. (75)] The SSIM data range L is computed per sample from the reference field. If a frame is nearly constant, L can be near zero, making the SSIM constants unstable. Consider using a dataset-global range or an explicit floor.

Circularity Check

1 steps flagged

MRD is partially circular with the TIC objective; core claims otherwise remain independent.

specific steps
  1. fitted input called prediction [Sec. 4.2 Eq. (20); App. C.2 Eq. (47); App. F.3 Eqs. (77)-(80); Sec. 5 Metrics]
    "LTIC = W−1 X w=1 κw Ex,τ,τM ,ϵ h ||∆xτ,τ+w −∆ˆxτ,τ+w ||2 2 i ... App. C.2: E h ||X τ+w −X τ ||2 2 i = 2 tr(Γ(0)−Γ(w)). App. F.3: 'we estimate the mixing rate λ by fitting the exponential decay model C(ℓ)≈exp(−λℓ), ℓ=0,...,K'; main text: 'the mixing rate is defined as the decay exponent of the full-trajectory autocorrelation curve, estimated from the first 20 frames'."

    TIC is trained to match finite-lag increments for lags w=1..5 (Eq. 20; Table 4). Appendix C.2 shows that, for a second-order stationary process, the expected squared increment equals 2tr(Γ(0)−Γ(w)), i.e., it is the same information as autocovariance decay. MRD (App. F.3) is estimated by fitting exp(−λℓ) to the normalized autocorrelation over the first 20 lags, and is then reported as evidence of 'mixing-rate-related dynamics'. Thus the MRD improvement is to a large degree a re-measurement of the quantity TIC directly minimizes, rather than an independent confirmation that full-trajectory temporal correlations are preserved. The short-lag overlap (w=1..5 trained vs. ℓ=0..20 measured) makes the circularity partial; PSDD, short-term RL2/SSIM, and inference-speed claims remain independent.

full rationale

The paper is not built on a load-bearing self-citation chain: the pixel MeanFlow foundation is cited to external authors [32], and the proposed contributions (WinC-MF masking, TIC, 1-NFE blockwise rollout) are evaluated against independent neural-operator baselines and a TIC ablation (Table 9). The main circularity I can exhibit is metric-level: TIC penalizes exactly the finite-lag increment statistics that Appendix C.2 equates to autocovariance decay, and MRD is estimated from that autocovariance decay, so the reported MRD gains are partly a re-statement of the training objective rather than an independent long-horizon validation. This affects only one of the five headline metrics; PSDD, TKED (through its dependence on fluctuation variance), RL2/SSIM, and the one-step speed advantage retain independent content. The remaining concern that matching lags 1..5 does not prove control of the full invariant measure or of 320/625-frame statistics is a correctness/evidence gap, not circularity, so I do not raise the score further.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The method introduces no new physical entities. Its predictive content rests on hand-chosen TIC lag weights, a masking rate, a window length, and the assumption that short-lag increment matching controls long-horizon statistics. These are reasonable modeling choices, but they are not derived from first principles.

free parameters (3)
  • TIC lag weights kappa_w = (0.4, 0.5, 0.8, 1.1, 1.2)
    Table 4; hand-selected to upweight longer lags; directly controls the TIC loss contribution in Eq. 20.
  • Masking rate upsilon = 0.8
    Table 4; controls the expected number of revealed frames (about two) and hence the task difficulty and stochasticity in WinC-MF.
  • Window length W = 6 (implied)
    Not stated as a single hyperparameter, but five TIC lag weights imply W=6; it sets context/block geometry and the maximum lag used in training.
axioms (4)
  • domain assumption Masked future frames leave irreducible uncertainty, so the forecasting task is not a deterministic copy.
    Appendix C.1; needed for WinC-MF to act as a probabilistic generator. Plausible for chaotic and projected dynamics, but not proven for these datasets.
  • domain assumption Trajectories are second-order stationary in time, so E||X_{tau+w}-X_tau||^2 = 2tr(Gamma(0)-Gamma(w)).
    Appendix C.2 Eqs. 46-47; used to justify TIC as a covariance/mixing regularizer. Approximately true for developed turbulence, but not verified on the evaluation windows.
  • ad hoc to paper Short-lag TIC (w <= 5) transfers to full-trajectory statistics and invariant measures.
    Core premise of the paper; only five lags are trained, while evaluations extend to 320 or 625 frames. No theorem guarantees this transfer; it is the main empirical hope of the method.
  • domain assumption In the projected TCF192 observable, finite-lag increments capture Mori-Zwanzig memory and unresolved contributions.
    Appendix C.2; motivates TIC for the non-Markovian slice dynamics. The argument is qualitative and no quantitative bound is given.

pith-pipeline@v1.3.0-alltime-deepseek · 27829 in / 14656 out tokens · 154708 ms · 2026-08-02T14:43:36.116310+00:00 · methodology

0 comments
read the original abstract

Fast surrogate modeling for high-dimensional physical dynamics requires more than low short-term error: useful models must roll out efficiently while preserving the statistical structure of long trajectories. Neural operators provide inexpensive autoregressive forecasts but can drift in turbulent regimes, whereas rolling diffusion and latent generative surrogates can represent stochastic transitions at the cost of multi-step denoising, noise-schedule design, or auxiliary compression models. We propose MeanFlow Long-term Invariant Spatiotemporal Consistency Autoregressive Models (MeLISA), a latent-free autoregressive generative surrogate built on pixel-space MeanFlow. MeLISA defines a blockwise stochastic transition kernel that generates each forecast block with a single model evaluation, avoiding latent encoders and iterative diffusion solvers at inference time. To stabilize long-horizon rollouts, MeLISA combines a Window-Consistency MeanFlow objective that learns conditional spatiotemporal generation from partially observed temporal windows with a Time Increment Consistency loss that constrains multi-lag finite increments and targets temporal-correlation structure. We evaluate MeLISA with compact UNet and scalable DiT backbones on two high-resolution benchmarks, extended 2D Kolmogorov flow at $256 \times 256$ and turbulent channel-flow slice at $192 \times 192$. MeLISA outperforms neural-operator baselines on short-term forecasting accuracy and long-horizon statistical metrics, including energy spectra, turbulent kinetic energy, and mixing-rate-related dynamics, while achieving inference speeds comparable to, and in some cases faster than, neural operators. To our knowledge, this is the first method for high-resolution one-step generation for physical dynamical systems with performance comparable to state-of-the-art deterministic surrogates.

Figures

Figures reproduced from arXiv: 2605.05540 by Tianyue Yang, Xiao Xue.

Figure 1
Figure 1. Figure 1: This diagram illustrates the two key mechanisms of MeanFlow Long-term Invariant [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Rollout results on an uncurated test trajectory from [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: (a) Radially averaged energy spectra for KF256 and TCF192. The neural-operator baselines [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: , computed over the first 40 frames as in the main text. Most models remain stable with 1–3 guidance frames but deteriorate sharply when 4 or 5 frames are provided. In contrast, MeLISA-∆-M and MeLISA-∆-B remain stable across the entire range. This behavior is consistent with prior observations that pixel-space self-supervised objectives typically require sufficiently large models to be effective [73, 78, 7… view at source ↗
Figure 5
Figure 5. Figure 5: Long-term PSDD loss against number of MeLISA guidance frames. The y-axis is in log [PITH_FULL_IMAGE:figures/full_fig_p028_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: TKED loss against number of MeLISA guidance frames. The y-axis is in log scale. [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Autoregressive rollout results for MeLISA- [PITH_FULL_IMAGE:figures/full_fig_p029_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Accumulation of RL2 error over the first 40 rollout frames on KF256. Although the spatial error remains visually reasonable, the spectral error grows much more severely during rollout ( [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Accumulation of PSDD over autoregressive rollouts on KF256, evaluated over the full [PITH_FULL_IMAGE:figures/full_fig_p030_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Stress test of maximum rollout length for MeLISA models. White frames indicate either [PITH_FULL_IMAGE:figures/full_fig_p031_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Evolution of CRPS on KF256. The upper panel shows short-term CRPS, and the lower [PITH_FULL_IMAGE:figures/full_fig_p032_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Trajectory ensemble generated by MeLISA- [PITH_FULL_IMAGE:figures/full_fig_p033_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Rollout results on an uncurated test trajectory from [PITH_FULL_IMAGE:figures/full_fig_p035_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Autocorrelation on the KF256 dataset. The upper panel shows the first 40 frames, while [PITH_FULL_IMAGE:figures/full_fig_p036_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Autocorrelation on the TCF192 dataset. The upper panel shows the first 40 frames, while [PITH_FULL_IMAGE:figures/full_fig_p037_15.png] view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

84 extracted references · 33 linked inside Pith

  1. [1]

    American mathematical society, 2022

    Lawrence C Evans.Partial differential equations, volume 19. American mathematical society, 2022

    2022

  2. [2]

    Direct numerical simulation of turbulent channel flow up to.Journal of fluid mechanics, 774:395–415, 2015

    Myoungkyu Lee and Robert D Moser. Direct numerical simulation of turbulent channel flow up to.Journal of fluid mechanics, 774:395–415, 2015

    2015

  3. [3]

    SIAM, 2004

    John C Strikwerda.Finite difference schemes and partial differential equations. SIAM, 2004

    2004

  4. [4]

    Finite volume methods.Handbook of numerical analysis, 7:713–1018, 2000

    Robert Eymard, Thierry Gallouët, and Raphaèle Herbin. Finite volume methods.Handbook of numerical analysis, 7:713–1018, 2000

    2000

  5. [5]

    An introduction to the finite element method.New York, 27(14), 1993

    Junuthula Narasimha Reddy. An introduction to the finite element method.New York, 27(14), 1993

    1993

  6. [6]

    Reynolds-averaged navier–stokes equations for turbulence modeling

    Giancarlo Alfonsi. Reynolds-averaged navier–stokes equations for turbulence modeling. 2009

    2009

  7. [7]

    Large-eddy simulation: achievements and challenges.Progress in aerospace sciences, 35(4):335–362, 1999

    Ugo Piomelli. Large-eddy simulation: achievements and challenges.Progress in aerospace sciences, 35(4):335–362, 1999

    1999

  8. [8]

    Prediff: Precipitation nowcasting with latent diffusion models

    Zhihan Gao, Xingjian Shi, Boran Han, Hao Wang, Xiaoyong Jin, Danielle Maddix, Yi Zhu, Mu Li, and Yuyang Bernie Wang. Prediff: Precipitation nowcasting with latent diffusion models. Advances in Neural Information Processing Systems, 36:78621–78656, 2023

    2023

  9. [9]

    Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

    Kaifeng Bi, Lingxi Xie, Hengheng Zhang, Xin Chen, Xiaotao Gu, and Qi Tian. Pangu-weather: A 3d high-resolution model for fast and accurate global weather forecast.arXiv preprint arXiv:2211.02556, 2022

    Pith/arXiv arXiv 2022

  10. [10]

    A self-attention ansatz for ab-initio quantum chemistry.arXiv preprint arXiv:2211.13672, 2022

    Ingrid von Glehn, James S Spencer, and David Pfau. A self-attention ansatz for ab-initio quantum chemistry.arXiv preprint arXiv:2211.13672, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Ferminets: Learning generative machines to generate efficient neural networks via generative synthesis

    Alexander Wong, Mohammad Javad Shafiee, Brendan Chwyl, and Francis Li. Ferminets: Learning generative machines to generate efficient neural networks via generative synthesis. arXiv preprint arXiv:1809.05989, 2018

    Pith/arXiv arXiv 2018

  12. [12]

    Crystalflow: a flow-based generative model for crystalline materials

    Xiaoshan Luo, Zhenyu Wang, Qingchang Wang, Xuechen Shao, Jian Lv, Lei Wang, Yanchao Wang, and Yanming Ma. Crystalflow: a flow-based generative model for crystalline materials. Nature Communications, 16(1):9267, 2025

    2025

  13. [13]

    Uni-flow: a unified autoregressive- diffusion model for complex multiscale flows.arXiv preprint arXiv:2602.15592, 2026

    Xiao Xue, Tianyue Yang, Mingyang Gao, Leyu Pan, Maida Wang, Kewei Zhu, Shuo Wang, Jiuling Li, Marco FP ten Eikelder, and Peter V Coveney. Uni-flow: a unified autoregressive- diffusion model for complex multiscale flows.arXiv preprint arXiv:2602.15592, 2026

    arXiv 2026

  14. [14]

    Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial differen- tial equations.arXiv preprint arXiv:2010.08895, 2020

    Pith/arXiv arXiv 2010

  15. [15]

    Learning nonlinear operators via deeponet based on the universal approximation theorem of operators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learning nonlinear operators via deeponet based on the universal approximation theorem of operators. Nature machine intelligence, 3(3):218–229, 2021

    2021

  16. [16]

    U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    Md Ashiqur Rahman, Zachary E Ross, and Kamyar Azizzadenesheli. U-no: U-shaped neural operators.arXiv preprint arXiv:2204.11127, 2022

    Pith/arXiv arXiv 2022

  17. [17]

    Neural operators with localized integral and differential kernels.arXiv preprint arXiv:2402.16845, 2024

    Miguel Liu-Schiaffini, Julius Berner, Boris Bonev, Thorsten Kurth, Kamyar Azizzadenesheli, and Anima Anandkumar. Neural operators with localized integral and differential kernels.arXiv preprint arXiv:2402.16845, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Towards stability of autoregressive neural operators.arXiv preprint arXiv:2306.10619, 2023

    Michael McCabe, Peter Harrington, Shashank Subramanian, and Jed Brown. Towards stability of autoregressive neural operators.arXiv preprint arXiv:2306.10619, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Training neural operators to preserve invariant measures of chaotic attractors.Advances in Neural Information Processing Systems, 36, 2024

    Ruoxi Jiang, Peter Y Lu, Elena Orlova, and Rebecca Willett. Training neural operators to preserve invariant measures of chaotic attractors.Advances in Neural Information Processing Systems, 36, 2024. 10

    2024

  20. [20]

    Integrating neural operators with diffusion models improves spectral representation in turbulence modeling.arXiv preprint arXiv:2409.08477, 2024

    Vivek Oommen, Aniruddha Bora, Zhen Zhang, and George Em Karniadakis. Integrating neural operators with diffusion models improves spectral representation in turbulence modeling.arXiv preprint arXiv:2409.08477, 2024

    Pith/arXiv arXiv 2024

  21. [21]

    Miti- gating spectral bias in neural operators via high-frequency scaling for physical systems.arXiv preprint arXiv:2503.13695, 2025

    Siavash Khodakarami, Vivek Oommen, Aniruddha Bora, and George Em Karniadakis. Miti- gating spectral bias in neural operators via high-frequency scaling for physical systems.arXiv preprint arXiv:2503.13695, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    2020

  23. [23]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Pith/arXiv arXiv 2011

  24. [24]

    Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  25. [25]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. 2023

    2023

  26. [26]

    Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models.arXiv preprint arXiv:2410.11081, 2024

    Pith/arXiv arXiv 2024

  27. [27]

    Benchmarking autoregressive conditional diffusion models for turbulent flow simulation.Neural Networks, page 108641, 2026

    Georg Kohl, Li-Wei Chen, and Nils Thuerey. Benchmarking autoregressive conditional diffusion models for turbulent flow simulation.Neural Networks, page 108641, 2026

    2026

  28. [28]

    Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

    Salva Rühling Cachay, Bo Zhao, Hailey Joren, and Rose Yu. Dyffusion: A dynamics-informed diffusion model for spatiotemporal forecasting.Advances in neural information processing systems, 36:45259–45287, 2023

    2023

  29. [29]

    Elucidated rolling diffusion models for probabilistic forecasting of complex dynamics.arXiv preprint arXiv:2506.20024, 2025

    Salva Rühling Cachay, Miika Aittala, Karsten Kreis, Noah Brenowitz, Arash Vahdat, Morteza Mardani, and Rose Yu. Elucidated rolling diffusion models for probabilistic forecasting of complex dynamics.arXiv preprint arXiv:2506.20024, 2025

    arXiv 2025

  30. [30]

    Probabilistic weather forecasting with machine learning.Nature, 637(8044):84–90, 2025

    Ilan Price, Alvaro Sanchez-Gonzalez, Ferran Alet, Tom R Andersson, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, Shakir Mohamed, Peter Battaglia, et al. Probabilistic weather forecasting with machine learning.Nature, 637(8044):84–90, 2025

    2025

  31. [31]

    Rolling diffusion models

    David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. arXiv preprint arXiv:2402.09470, 2024

    Pith/arXiv arXiv 2024

  32. [32]

    One-step latent-free image generation with pixel mean flows.arXiv preprint arXiv:2601.22158, 2026

    Yiyang Lu, Susie Lu, Qiao Sun, Hanhong Zhao, Zhicheng Jiang, Xianbang Wang, Tianhong Li, Zhengyang Geng, and Kaiming He. One-step latent-free image generation with pixel mean flows.arXiv preprint arXiv:2601.22158, 2026

    Pith/arXiv arXiv 2026

  33. [33]

    Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    Pan Du, Meet Hemant Parikh, Xiantao Fan, Xin-Yang Liu, and Jian-Xun Wang. Conditional neural field latent diffusion model for generating spatiotemporal turbulence.Nature Communi- cations, 15(1):10416, 2024

    2024

  34. [34]

    Elucidating the design choice of probability paths in flow matching for forecasting.arXiv preprint arXiv:2410.03229, 2024

    Soon Hoe Lim, Yijin Wang, Annan Yu, Emma Hart, Michael W Mahoney, Xiaoye S Li, and N Benjamin Erichson. Elucidating the design choice of probability paths in flow matching for forecasting.arXiv preprint arXiv:2410.03229, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6038–6047, 2023

    2023

  36. [36]

    A physics-informed diffusion model for high- fidelity flow field reconstruction.Journal of Computational Physics, 478:111972, 2023

    Dule Shu, Zijie Li, and Amir Barati Farimani. A physics-informed diffusion model for high- fidelity flow field reconstruction.Journal of Computational Physics, 478:111972, 2023

    2023

  37. [37]

    Thomas Y . L. Lin, Jiachen Yao, Lufang Chiang, Julius Berner, and Anima Anandkumar. Decoupled diffusion sampling for inverse problems on function spaces, 2026. URL https: //arxiv.org/abs/2601.23280. 11

    arXiv 2026

  38. [38]

    Meno: Meanflow-enhanced neural operators for dynamical systems, 2026

    Tianyue Yang and Xiao Xue. Meno: Meanflow-enhanced neural operators for dynamical systems, 2026. URLhttps://arxiv.org/abs/2604.06881

    Pith/arXiv arXiv 2026

  39. [39]

    Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35: 26565–26577, 2022

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models.Advances in neural information processing systems, 35: 26565–26577, 2022

    2022

  40. [40]

    Generative ai for fast and accurate statistical computation of fluids.arXiv preprint arXiv:2409.18359, 2024

    Roberto Molinaro, Samuel Lanthaler, Bogdan Raoni ´c, Tobias Rohner, Victor Armegioiu, Stephan Simonis, Dana Grund, Yannick Ramic, Zhong Yi Wan, Fei Sha, et al. Generative ai for fast and accurate statistical computation of fluids.arXiv preprint arXiv:2409.18359, 2024

    Pith/arXiv arXiv 2024

  41. [41]

    Inductive moment matching.arXiv preprint arXiv:2503.07565, 2025

    Linqi Zhou, Stefano Ermon, and Jiaming Song. Inductive moment matching.arXiv preprint arXiv:2503.07565, 2025

    Pith/arXiv arXiv 2025

  42. [42]

    One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024

    Kevin Frans, Danijar Hafner, Sergey Levine, and Pieter Abbeel. One step diffusion via shortcut models.arXiv preprint arXiv:2410.12557, 2024

    Pith/arXiv arXiv 2024

  43. [43]

    Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447, 2025

    Zhengyang Geng, Mingyang Deng, Xingjian Bai, J Zico Kolter, and Kaiming He. Mean flows for one-step generative modeling.arXiv preprint arXiv:2505.13447, 2025

    Pith/arXiv arXiv 2025

  44. [44]

    Improved mean flows: On the challenges of fastforward generative models.arXiv preprint arXiv:2512.02012, 2025

    Zhengyang Geng, Yiyang Lu, Zongze Wu, Eli Shechtman, J Zico Kolter, and Kaiming He. Improved mean flows: On the challenges of fastforward generative models.arXiv preprint arXiv:2512.02012, 2025

    Pith/arXiv arXiv 2025

  45. [45]

    Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    Pith/arXiv arXiv 2023

  46. [46]

    Osv: One step is enough for high-quality image to video generation

    Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Jiangning Zhang, Hao Chen, Mingmin Chi, Yabiao Wang, and Wenhan Luo. Osv: One step is enough for high-quality image to video generation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12585–12594, 2025

    2025

  47. [47]

    Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109, 2023

    Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model.arXiv preprint arXiv:2312.09109, 2023

    Pith/arXiv arXiv 2023

  48. [48]

    Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

    Qiang Liu. Rectified flow: A marginal preserving approach to optimal transport.arXiv preprint arXiv:2209.14577, 2022

    Pith/arXiv arXiv 2022

  49. [49]

    Back to basics: Let denoising generative models denoise, 2026

    Tianhong Li and Kaiming He. Back to basics: Let denoising generative models denoise, 2026. URLhttps://arxiv.org/abs/2511.13720

    Pith/arXiv arXiv 2026

  50. [50]

    Walrus: A cross-domain foundation model for continuum dynamics.arXiv preprint arXiv:2511.15684, 2025

    Michael McCabe, Payel Mukhopadhyay, Tanya Marwah, Bruno Regaldo-Saint Blancard, Fran- cois Rozet, Cristiana Diaconu, Lucas Meyer, Kaze WK Wong, Hadi Sotoudeh, Alberto Bietti, et al. Walrus: A cross-domain foundation model for continuum dynamics.arXiv preprint arXiv:2511.15684, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational conference on machine learning, pages 8162–8171. PMLR, 2021

    2021

  52. [52]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 4195–4205, 2023

    2023

  53. [53]

    Muon is scalable for llm training.arXiv preprint arXiv:2502.16982, 2025

    Jingyuan Liu, Jianlin Su, Xingcheng Yao, Zhejun Jiang, Guokun Lai, Yulun Du, Yidao Qin, Weixin Xu, Enzhe Lu, Junjie Yan, et al. Muon is scalable for llm training.arXiv preprint arXiv:2502.16982, 2025

    Pith/arXiv arXiv 2025

  54. [54]

    Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4): 600–612, 2004

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing, 13(4): 600–612, 2004

    2004

  55. [55]

    Image quality metrics: Psnr vs

    Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In2010 20th international conference on pattern recognition, pages 2366–2369. IEEE, 2010. 12

    2010

  56. [56]

    Generalized teacher forcing for learning chaotic dynamics.arXiv preprint arXiv:2306.04406, 2023

    Florian Hess, Zahra Monfared, Manuel Brenner, and Daniel Durstewitz. Generalized teacher forcing for learning chaotic dynamics.arXiv preprint arXiv:2306.04406, 2023

    Pith/arXiv arXiv 2023

  57. [57]

    A scalable generative model for dynamical system reconstruction from neuroimaging data.Advances in Neural Information Processing Systems, 37:80328–80362, 2024

    Eric V olkmann, Alena Brändle, Daniel Durstewitz, and Georgia Koppe. A scalable generative model for dynamical system reconstruction from neuroimaging data.Advances in Neural Information Processing Systems, 37:80328–80362, 2024

    2024

  58. [58]

    Flax: A neural network library and ecosystem for JAX, 2024

    Jonathan Heek, Anselm Levskaya, Avital Oliver, Marvin Ritter, Bertrand Rondepierre, Andreas Steiner, and Marc van Zee. Flax: A neural network library and ecosystem for JAX, 2024. URL http://github.com/google/flax

    2024

  59. [59]

    Shampoo: Preconditioned stochastic tensor optimization

    Vineet Gupta, Tomer Koren, and Yoram Singer. Shampoo: Preconditioned stochastic tensor optimization. InInternational Conference on Machine Learning, pages 1842–1850. PMLR, 2018

    2018

  60. [60]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

    Pith/arXiv arXiv 2014

  61. [61]

    Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Do- minik Lorenz, Yam Levi, Zion English, Vikram V oleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets.arXiv preprint arXiv:2311.15127, 2023

    Pith/arXiv arXiv 2023

  62. [62]

    Diffusion probabilistic modeling for video generation.Entropy, 25(10):1469, 2023

    Ruihan Yang, Prakhar Srivastava, and Stephan Mandt. Diffusion probabilistic modeling for video generation.Entropy, 25(10):1469, 2023

    2023

  63. [63]

    Fifo-diffusion: Generating infinite videos from text without training.Advances in Neural Information Processing Systems, 37: 89834–89868, 2024

    Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training.Advances in Neural Information Processing Systems, 37: 89834–89868, 2024

    2024

  64. [64]

    Seine: Short-to-long video diffusion model for generative transition and prediction

    Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. InThe Twelfth International Conference on Learning Representations, 2023

    2023

  65. [65]

    Ar-diffusion: Auto-regressive diffusion model for text generation.Advances in Neural Information Processing Systems, 36:39957–39974, 2023

    Tong Wu, Zhihao Fan, Xiao Liu, Hai-Tao Zheng, Yeyun Gong, Jian Jiao, Juntao Li, Jian Guo, Nan Duan, Weizhu Chen, et al. Ar-diffusion: Auto-regressive diffusion model for text generation.Advances in Neural Information Processing Systems, 36:39957–39974, 2023

    2023

  66. [66]

    Rflav: Rolling flow matching for infinite audio video generation.arXiv preprint arXiv:2503.08307, 2025

    Alex Ergasti, Giuseppe Gabriele Tarollo, Filippo Botti, Tomaso Fontanini, Claudio Ferrari, Massimo Bertozzi, and Andrea Prati. Rflav: Rolling flow matching for infinite audio video generation.arXiv preprint arXiv:2503.08307, 2025

    Pith/arXiv arXiv 2025

  67. [67]

    Probabilistic forecasting via autoregressive flow matching.arXiv preprint arXiv:2503.10375, 2025

    Ahmed ElGazzar and Marcel van Gerven. Probabilistic forecasting via autoregressive flow matching.arXiv preprint arXiv:2503.10375, 2025

    arXiv 2025

  68. [68]

    Rectified flows for fast multiscale fluid flow modeling.arXiv preprint arXiv:2506.03111, 2025

    Victor Armegioiu, Yannick Ramic, and Siddhartha Mishra. Rectified flows for fast multiscale fluid flow modeling.arXiv preprint arXiv:2506.03111, 2025

    arXiv 2025

  69. [69]

    Deep generative learning via schrödinger bridge

    Gefei Wang, Yuling Jiao, Qian Xu, Yang Wang, and Can Yang. Deep generative learning via schrödinger bridge. InInternational conference on machine learning, pages 10794–10804. PMLR, 2021

    2021

  70. [70]

    On the design of one-step diffusion via shortcutting flow paths.arXiv preprint arXiv:2512.11831, 2025

    Haitao Lin, Peiyan Hu, Minsi Ren, Zhifeng Gao, Zhi-Ming Ma, Tailin Wu, Stan Z Li, et al. On the design of one-step diffusion via shortcutting flow paths.arXiv preprint arXiv:2512.11831, 2025

    arXiv 2025

  71. [71]

    Alphaflow: Understanding and improving meanflow models

    Huijie Zhang, Aliaksandr Siarohin, Willi Menapace, Michael Vasilkovsky, Sergey Tulyakov, Qing Qu, and Ivan Skorokhodov. Alphaflow: Understanding and improving meanflow models. arXiv preprint arXiv:2510.20771, 2025

    arXiv 2025

  72. [72]

    simple diffusion: End-to-end diffusion for high resolution images

    Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. InInternational Conference on Machine Learning, pages 13213– 13232. PMLR, 2023. 13

    2023

  73. [73]

    Masked au- toencoders are scalable vision learners, 2021

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked au- toencoders are scalable vision learners, 2021. URL https://arxiv.org/abs/2111.06377

    Pith/arXiv arXiv 2021

  74. [74]

    Masked trajectory models for prediction, representation, and control,

    Philipp Wu, Arjun Majumdar, Kevin Stone, Yixin Lin, Igor Mordatch, Pieter Abbeel, and Aravind Rajeswaran. Masked trajectory models for prediction, representation, and control,

  75. [75]

    Memory effects in irreversible thermodynamics.Physical Review, 124(4):983, 1961

    Robert Zwanzig. Memory effects in irreversible thermodynamics.Physical Review, 124(4):983, 1961

    1961

  76. [76]

    Lattice boltzmann method for fluid flows.Annual review of fluid mechanics, 30(1):329–364, 1998

    Shiyi Chen and Gary D Doolen. Lattice boltzmann method for fluid flows.Annual review of fluid mechanics, 30(1):329–364, 1998

    1998

  77. [77]

    Oxford University Press, 2001

    Sauro Succi.The Lattice Boltzmann Equation for Fluid Dynamics and Beyond. Oxford University Press, 2001

    2001

  78. [78]

    Self-supervised learning from images with a joint- embedding predictive architecture

    Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint- embedding predictive architecture. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15619–15629, 2023

    2023

  79. [79]

    Emerging opportuni- ties and challenges for the future of reservoir computing.Nature Communications, 15(1):2056, 2024

    Min Yan, Can Huang, Peter Bienstman, Peter Tino, Wei Lin, and Jie Sun. Emerging opportuni- ties and challenges for the future of reservoir computing.Nature Communications, 15(1):2056, 2024

    2056

  80. [80]

    Straight velocity boundaries in the lattice boltzmann method.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 77(5):056703, 2008

    Jonas Latt, Bastien Chopard, Orestis Malaspinas, Michel Deville, and Andreas Michler. Straight velocity boundaries in the lattice boltzmann method.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 77(5):056703, 2008

    2008

Showing first 80 references.