pith. sign in

arxiv: 2605.26535 · v1 · pith:HMK57WBXnew · submitted 2026-05-26 · 💻 cs.LG · cs.AI· cs.CV· cs.NA· math.NA

Recursive Flow Matching

Jiahe Huang , Sihan Xu , Sharvaree Vadgama , Rose Yu This is my paper

Pith reviewed 2026-06-29 19:18 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CVcs.NAmath.NA
keywords flow matchinggenerative modelsspatiotemporal dynamicsself-consistencydiscretization errorsscientific emulationphysics systems
0
0 comments X

The pith

Recursive Flow Matching enforces self-consistency across discretization scales to enable accurate one- and few-step generation of spatiotemporal dynamics.

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

The paper presents Recursive Flow Matching as a generative framework for forecasting complex physics systems and spatiotemporal dynamics. It enforces self-consistency to align trajectories at different discretization levels, which reduces the errors that typically arise in fast sampling methods. A sympathetic reader would care because this targets the long-standing speed-versus-accuracy trade-off in scientific emulation. If the approach holds, it would allow high-fidelity forecasts in one to four steps that match the quality of far slower multi-step solvers. The work reports concrete gains on scientific benchmarks, including substantial speedups over diffusion-based methods and lower error than standard flow matching.

Core claim

Recursive Flow Matching (RecFM) is a generative framework for forecasting complex spatiotemporal dynamics that enforces self-consistency to align trajectories across discretization scales. This alignment reduces discretization errors and improves performance across metrics for physics-based tasks. The method achieves high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems with performance comparable to state-of-the-art multi-step solvers, delivering up to a 20× speedup over leading diffusion-based emulators while improving predictive accuracy and reducing mean squared error by over 15% compared to vanilla flow matching.

What carries the argument

Recursive Flow Matching, the mechanism that enforces self-consistency to align trajectories across discretization scales in a flow-based generative model.

If this is right

  • Enables high-fidelity one- and few-step (2-4 step) generation for scientific systems that matches multi-step solver performance.
  • Delivers up to 20× speedup over diffusion-based emulators on challenging scientific benchmarks.
  • Reduces mean squared error by more than 15% relative to vanilla flow matching.
  • Offers a scalable path to real-time scientific emulation without sacrificing physical accuracy.

Where Pith is reading between the lines

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

  • The self-consistency idea could be tested in other flow- or diffusion-based models for dynamics to see if similar error reductions occur.
  • The method may extend to additional domains such as climate modeling or fluid dynamics where discretization errors are a known bottleneck.
  • Further work could examine whether the same alignment principle improves performance at even coarser discretization levels.

Load-bearing premise

Enforcing self-consistency to align trajectories across discretization scales will reduce discretization errors enough to produce the reported accuracy and speed gains on representative scientific tasks.

What would settle it

A controlled test on one of the paper's scientific benchmarks in which RecFM one- or four-step outputs show higher error than a standard multi-step solver or fail to reach the claimed speedup over diffusion emulators.

Figures

Figures reproduced from arXiv: 2605.26535 by Jiahe Huang, Rose Yu, Sharvaree Vadgama, Sihan Xu.

Figure 1
Figure 1. Figure 1: Comparison of flow matching paradigms. (a) Flow Matching (FM) learns a direct trajectory that transports samples from the data distribution (x0) to the noise distribution (x1). (b) Recursive Flow Matching (RecFM) augments this with recursively scaled trajectories (brown, blue, and red arrows) that intersect at shared spatial states (xt), enabling cross-scale trajectory alignment and consistency training al… view at source ↗
Figure 2
Figure 2. Figure 2: Pendulum trajec￾tories and velocities for the primary trajectory (v (1) , or￾ange) and attenuated trajec￾tories (v (i) , i > 1, blue). We draw inspiration from the recursive movement of an ideal1 wall￾bouncing pendulum to design our method, RecFM. Below, we introduce the pendulum model, followed by the secondary trajectory formulation and the updated loss function for RecFM. 3.1 Physics Intuition Let’s con… view at source ↗
Figure 3
Figure 3. Figure 3: Roll-out results of the Helmholtz Staircase equation. Visual comparison of Ground Truth against RecFM and VideoPDE (best-performed baseline) for two channels, with the bottom rows indicating absolute errors. Columns correspond to dataset timesteps. The variation observed at Step 48 is displayed in an enlarged view on the right. CRPS(F, y) = Z ∞ −∞ (F(z) − 1[z ≥ y])2 dz (12) In practice, we use the unbiased… view at source ↗
Figure 4
Figure 4. Figure 4: Validation MSE versus NFE dur￾ing training. RecFM converges faster than the diffusion-based model VideoPDE and maintains consistently lower validation error. Training Stability. We measure training progress of Navier-Stokes Flow in terms of the number of function evaluations (NFE), defined as the total number of vector field evaluations (i.e., forward passes) during optimization. As shown in [PITH_FULL_IM… view at source ↗
Figure 5
Figure 5. Figure 5: MSE vs. inference steps on the Navier-Stokes benchmark. RecFM achieves optimal performance with one- and two-step generation, while increasing the number of steps leads to error accumulation. C.2 Influence of Recursion Depth D We study the effect of recursion depth D on model performance. The depth D controls the number of trajectory scales used during training, thereby governing the strength of multi-scal… view at source ↗
Figure 6
Figure 6. Figure 6: Flow matching validation MSE versus NFE during training. RecFM converges faster than Vanilla FM and maintains consistently lower validation error. D Architecture and Implementation Details We adopt the state-of-the-art Hierarchical Video Diffusion Transformer (HV-DiT) backbone from VideoPDE [35], with the sole modification that the input mask channel is removed. Unlike DYffusion [15], whose forecasting-ori… view at source ↗
Figure 7
Figure 7. Figure 7: More roll-out results of the Helmholtz Staircase equation. Visual comparison of Ground Truth against RecFM and VideoPDE (best-performed baseline) for two channels, with the bottom rows indicating absolute errors. Columns correspond to dataset timesteps. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Navier-Stokes rollout sample. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Average kinetic energy over time. ⟨E(t)⟩ normalized by ⟨Et=0 real ⟩. 100% corresponds to the initial ground-truth energy. 20 [PITH_FULL_IMAGE:figures/full_fig_p020_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Selected samples from our 256 × 256 resolution RecFM-XL model. I Recursive Flow Matching for Image Generation [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Uncurated 256 × 256 RecFM-XL samples. Each panel shows samples from a different ImageNet class. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Uncurated 256 × 256 RecFM-XL samples (Continued). Each panel shows samples from a different ImageNet class. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
read the original abstract

Generative models have emerged as a powerful paradigm for solving physics systems and modeling complex spatiotemporal dynamics. However, achieving high physical accuracy without incurring high computational cost remains a fundamental challenge, as existing approaches face a critical speed-fidelity trade-off. In this work, we introduce Recursive Flow Matching (RecFM), a generative framework for forecasting complex spatiotemporal dynamics. RecFM enforces self-consistency to align trajectories across discretization scales, reducing discretization errors and improving performance across metrics for physics-based tasks. To our knowledge, this is the first method to achieve high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems with performance comparable to state-of-the-art multi-step solvers. Across challenging scientific benchmarks, RecFM achieves up to a 20$\times$ speedup over leading diffusion-based emulators while improving predictive accuracy. Furthermore, RecFM reduces mean squared error by over 15% compared to vanilla flow matching, offering a scalable and efficient solution for real-time scientific emulation.

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

2 major / 0 minor

Summary. The manuscript introduces Recursive Flow Matching (RecFM), a generative framework for forecasting complex spatiotemporal dynamics. It claims that enforcing self-consistency to align trajectories across discretization scales reduces discretization errors, enabling high-fidelity one- and few-step (2-4 step) dynamic generation for scientific systems with performance comparable to state-of-the-art multi-step solvers. Reported results include up to 20× speedup over leading diffusion-based emulators and over 15% MSE reduction versus vanilla flow matching on challenging scientific benchmarks.

Significance. If the self-consistency mechanism is well-defined, mathematically justified, and empirically shown to reduce discretization error without introducing new biases, the work could meaningfully advance real-time scientific emulation by improving the speed-fidelity trade-off in generative models. The claim of being the first method to achieve such few-step performance on scientific tasks would be noteworthy if supported by detailed derivations, algorithms, and controlled experiments. However, the provided abstract supplies no equations, training objectives, or implementation details, preventing assessment of whether the central mechanism is novel or effective.

major comments (2)
  1. Abstract: the central claim that self-consistency enforcement reduces discretization errors and delivers the reported accuracy/speed gains cannot be evaluated because no mathematical formulation, loss term, algorithm, or pseudocode is supplied for how alignment across discretization scales is achieved or enforced.
  2. Abstract: without access to the training objective, vector-field parameterization, or benchmark protocols, it is impossible to determine whether the claimed >15% MSE reduction and 20× speedup arise from the self-consistency mechanism, from unmentioned architecture changes, or from benchmark selection.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their review of our manuscript on Recursive Flow Matching. We address the two major comments regarding the abstract below, clarifying that the full manuscript contains the requested mathematical details, algorithms, and experimental protocols.

read point-by-point responses

  1. Referee: Abstract: the central claim that self-consistency enforcement reduces discretization errors and delivers the reported accuracy/speed gains cannot be evaluated because no mathematical formulation, loss term, algorithm, or pseudocode is supplied for how alignment across discretization scales is achieved or enforced.

    Authors: Abstracts are conventionally kept equation-free for brevity and accessibility. The full manuscript defines the self-consistency mechanism in Section 2, presents the training objective (including the alignment loss across scales) in Section 3.1, and provides the recursive procedure with pseudocode in Algorithm 1. These sections directly support the central claim. revision: no

  2. Referee: Abstract: without access to the training objective, vector-field parameterization, or benchmark protocols, it is impossible to determine whether the claimed >15% MSE reduction and 20× speedup arise from the self-consistency mechanism, from unmentioned architecture changes, or from benchmark selection.

    Authors: Section 4 contains controlled experiments that isolate the self-consistency component via ablations against vanilla flow matching on identical architectures and benchmarks. The training objective appears in Equation (5), vector-field parameterization in Section 2.3, and full benchmark protocols (datasets, metrics, and evaluation) in Section 4.1. These establish that gains derive from the proposed mechanism. revision: no

Circularity Check

0 steps flagged

No circularity: claims rest on proposed self-consistency mechanism without reduction to inputs by construction

full rationale

The abstract and provided context introduce RecFM as a generative framework whose core innovation is enforcing self-consistency to align trajectories across discretization scales. No equations, parameter-fitting steps, self-citations, or uniqueness theorems are quoted that would make any performance claim (speedup, MSE reduction) equivalent to the inputs by definition. The derivation chain is therefore self-contained: the method is presented as a novel construction whose benefits are asserted to follow from the (unstated) consistency enforcement rather than from renaming or refitting existing quantities.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; full assessment impossible without manuscript.

pith-pipeline@v0.9.1-grok · 5708 in / 1161 out tokens · 39323 ms · 2026-06-29T19:18:26.152872+00:00 · methodology

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

54 extracted references · 22 canonical work pages · 8 internal anchors

  1. [1]

    John Wiley & Sons, 2012

    Gouri Dhatt, Emmanuel Lefrançois, and Gilbert Touzot.Finite element method. John Wiley & Sons, 2012

    2012

  2. [2]

    Nektar++: An open-source spectral/hp element framework.Computer physics communications, 192:205–219, 2015

    Chris D Cantwell, David Moxey, Andrew Comerford, Alessandro Bolis, Gabriele Rocco, Gianmarco Mengaldo, Daniele De Grazia, Sergey Yakovlev, J-E Lombard, Dirk Ekelschot, et al. Nektar++: An open-source spectral/hp element framework.Computer physics communications, 192:205–219, 2015

    2015

  3. [3]

    Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    Nikola Kovachki, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes.Journal of Machine Learning Research, 24(89):1–97, 2023

    2023

  4. [4]

    Fourier Neural Operator for Parametric Partial Differential Equations

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  5. [5]

    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

  6. [6]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

    2019

  7. [7]

    Multifidelity modeling for physics-informed neural networks (pinns).Journal of Computational Physics, 451:110844, 2022

    Michael Penwarden, Shandian Zhe, Akil Narayan, and Robert M Kirby. Multifidelity modeling for physics-informed neural networks (pinns).Journal of Computational Physics, 451:110844, 2022

    2022

  8. [8]

    Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, 630(8016):493–500, 2024

    Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with alphafold 3.Nature, 630(8016):493–500, 2024

    2024

  9. [9]

    Simul- taneous modeling of protein conformation and dynamics via autoregression

    Yuning Shen, Lihao Wang, Huizhuo Yuan, Yan Wang, Bangji Yang, and Quanquan Gu. Simul- taneous modeling of protein conformation and dynamics via autoregression. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  10. [10]

    A generative model for inorganic materials design.Nature, 639(8055):624–632, 2025

    Claudio Zeni, Robert Pinsler, Daniel Zügner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabbé, Shoko Ueda, et al. A generative model for inorganic materials design.Nature, 639(8055):624–632, 2025

    2025

  11. [11]

    Samudrace: Fast and accurate coupled climate modeling with 3d ocean and atmosphere emulators.arXiv preprint arXiv:2509.12490, 2025

    James PC Duncan, Elynn Wu, Surya Dheeshjith, Adam Subel, Troy Arcomano, Spencer K Clark, Brian Henn, Anna Kwa, Jeremy McGibbon, W Andre Perkins, et al. Samudrace: Fast and accurate coupled climate modeling with 3d ocean and atmosphere emulators.arXiv preprint arXiv:2509.12490, 2025

    work page arXiv 2025

  12. [12]

    Ace2: accurately learning sub- seasonal to decadal atmospheric variability and forced responses.npj Climate and Atmospheric Science, 8(1):205, 2025

    Oliver Watt-Meyer, Brian Henn, Jeremy McGibbon, Spencer K Clark, Anna Kwa, W Andre Perkins, Elynn Wu, Lucas Harris, and Christopher S Bretherton. Ace2: accurately learning sub- seasonal to decadal atmospheric variability and forced responses.npj Climate and Atmospheric Science, 8(1):205, 2025

    2025

  13. [13]

    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

    work page arXiv 2025

  14. [14]

    Diffusionpde: Generative pde-solving under partial observation.Advances in Neural Information Processing Systems, 37:130291–130323, 2024

    Jiahe Huang, Guandao Yang, Zichen Wang, and Jeong Joon Park. Diffusionpde: Generative pde-solving under partial observation.Advances in Neural Information Processing Systems, 37:130291–130323, 2024

    2024

  15. [15]

    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. 10

    2023

  16. [16]

    Yilin Zhuang, Sibo Cheng, and Karthik Duraisamy. Spatially-aware diffusion models with cross-attention for global field reconstruction with sparse observations.Computer Methods in Applied Mechanics and Engineering, 435:117623, 2025

    2025

  17. [17]

    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

  18. [18]

    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

  19. [19]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  20. [20]

    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

  21. [21]

    Riemannian continuous normalizing flows.Advances in neural information processing systems, 33:2503–2515, 2020

    Emile Mathieu and Maximilian Nickel. Riemannian continuous normalizing flows.Advances in neural information processing systems, 33:2503–2515, 2020

    2020

  22. [22]

    arXiv preprint arXiv:2302.03660 , year=

    Ricky TQ Chen and Yaron Lipman. Flow matching on general geometries.arXiv preprint arXiv:2302.03660, 2023

    work page arXiv 2023

  23. [23]

    Mean Flows for One-step Generative Modeling

    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

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Flow Matching for Generative Modeling

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

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  25. [25]

    Physics- constrained fine-tuning of flow-matching models for generation and inverse problems.arXiv preprint arXiv:2508.09156, 2025

    Jan Tauberschmidt, Sophie Fellenz, Sebastian J V ollmer, and Andrew B Duncan. Physics- constrained fine-tuning of flow-matching models for generation and inverse problems.arXiv preprint arXiv:2508.09156, 2025

    work page arXiv 2025

  26. [26]

    Cyclenet: Rethinking cycle consistency in text-guided diffusion for image manipulation.Advances in Neural Information Processing Systems, 36:10359–10384, 2023

    Sihan Xu, Ziqiao Ma, Yidong Huang, Honglak Lee, and Joyce Chai. Cyclenet: Rethinking cycle consistency in text-guided diffusion for image manipulation.Advances in Neural Information Processing Systems, 36:10359–10384, 2023

    2023

  27. [27]

    One Step Diffusion via Shortcut Models

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

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    Multi-student diffusion distillation for better one-step generators.arXiv preprint arXiv:2410.23274, 2024

    Yanke Song, Jonathan Lorraine, Weili Nie, Karsten Kreis, and James Lucas. Multi-student diffusion distillation for better one-step generators.arXiv preprint arXiv:2410.23274, 2024

    work page arXiv 2024

  29. [29]

    On understanding and overcoming spectral biases of deep neural network learning methods for solving pdes.Journal of Computational Physics, 530:113905, 2025

    Zhi-Qin John Xu, Lulu Zhang, and Wei Cai. On understanding and overcoming spectral biases of deep neural network learning methods for solving pdes.Journal of Computational Physics, 530:113905, 2025

    2025

  30. [30]

    Trajectory flow matching with applications to clinical time series modelling

    Xi Zhang, Yuan Pu, Yuki Kawamura, Andrew Loza, Yoshua Bengio, Dennis L Shung, and Alexander Tong. Trajectory flow matching with applications to clinical time series modelling. Advances in Neural Information Processing Systems, 37:107198–107224, 2024

    2024

  31. [31]

    arXiv preprint arXiv:2510.03569 , year=

    Mohammad Mohaiminul Islam, Thijs P Kuipers, Sharvaree Vadgama, Coen de Vente, Afsana Khan, Clara I Sánchez, and Erik J Bekkers. Longitudinal flow matching for trajectory modeling. arXiv preprint arXiv:2510.03569, 2025

    work page arXiv 2025

  32. [32]

    Inversion-free image editing with natural language.arXiv preprint arXiv:2312.04965, 2023

    Sihan Xu, Yidong Huang, Jiayi Pan, Ziqiao Ma, and Joyce Chai. Inversion-free image editing with natural language.arXiv preprint arXiv:2312.04965, 2023

    work page arXiv 2023

  33. [33]

    Space-time continuous pde forecasting using equivariant neural fields.Advances in Neural Information Processing Systems, 37:76553–76577, 2024

    David M Knigge, David R Wessels, Riccardo Valperga, Samuele Papa, Jan-Jakob Sonke, Efstratios Gavves, and Erik J Bekkers. Space-time continuous pde forecasting using equivariant neural fields.Advances in Neural Information Processing Systems, 37:76553–76577, 2024

    2024

  34. [34]

    Guided diffusion sampling on function spaces with applications to pdes.arXiv preprint arXiv:2505.17004, 2025

    Jiachen Yao, Abbas Mammadov, Julius Berner, Gavin Kerrigan, Jong Chul Ye, Kamyar Az- izzadenesheli, and Anima Anandkumar. Guided diffusion sampling on function spaces with applications to pdes.arXiv preprint arXiv:2505.17004, 2025. 11

    work page arXiv 2025

  35. [35]

    Videopde: Unified generative pde solving via video inpainting diffusion models.arXiv preprint arXiv:2506.13754, 2025

    Edward Li, Zichen Wang, Jiahe Huang, and Jeong Joon Park. Videopde: Unified generative pde solving via video inpainting diffusion models.arXiv preprint arXiv:2506.13754, 2025

    work page arXiv 2025

  36. [36]

    Mcvd-masked conditional video dif- fusion for prediction, generation, and interpolation.Advances in neural information processing systems, 35:23371–23385, 2022

    Vikram V oleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video dif- fusion for prediction, generation, and interpolation.Advances in neural information processing systems, 35:23371–23385, 2022

    2022

  37. [37]

    WIND: Weather Inverse Diffusion for Zero-Shot Atmospheric Modeling

    Michael Aich, Andreas Fürst, Florian Sestak, Carlos Ruiz-Gonzalez, Niklas Boers, and Johannes Brandstetter. Wind: Weather inverse diffusion for zero-shot atmospheric modeling.arXiv preprint arXiv:2602.03924, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  38. [38]

    The well: a large-scale collection of diverse physics simulations for machine learning.Advances in Neural Information Processing Systems, 37:44989–45037, 2024

    Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina J Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B Dalziel, Drummond B Fielding, et al. The well: a large-scale collection of diverse physics simulations for machine learning.Advances in Neural Information Processing Systems, 37:44989–45037, 2024

    2024

  39. [39]

    Rolling diffusion models.arXiv preprint arXiv:2402.09470, 2024

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

    work page arXiv 2024

  40. [40]

    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

  41. [41]

    Probing equivariance and symmetry breaking in convolu- tional networks.arXiv preprint arXiv:2501.01999, 2025

    Sharvaree Vadgama, Mohammad Mohaiminul Islam, Domas Buracas, Christian Shewmake, Artem Moskalev, and Erik Bekkers. Probing equivariance and symmetry breaking in convolu- tional networks.arXiv preprint arXiv:2501.01999, 2025

    work page arXiv 2025

  42. [42]

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

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow.arXiv preprint arXiv:2209.03003, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  43. [43]

    Generative Modeling via Drifting

    Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative modeling via drifting.arXiv preprint arXiv:2602.04770, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Generalised flow maps for few-step generative modelling on riemannian manifolds.arXiv preprint arXiv:2510.21608, 2025

    Oscar Davis, Michael S Albergo, Nicholas M Boffi, Michael M Bronstein, and Avishek Joey Bose. Generalised flow maps for few-step generative modelling on riemannian manifolds.arXiv preprint arXiv:2510.21608, 2025

    work page arXiv 2025

  45. [45]

    Physics vs distributions: Pareto optimal flow matching with physics constraints

    Giacomo Baldan, Qiang Liu, Alberto Guardone, and Nils Thuerey. Physics vs distributions: Pareto optimal flow matching with physics constraints. InThe Fourteenth International Confer- ence on Learning Representations, 2026

    2026

  46. [46]

    Improvements of the daily optimum interpolation sea surface temperature (doisst) version 2.1.Journal of Climate, 34(8):2923–2939, 2021

    Boyin Huang, Chunying Liu, Viva Banzon, Eric Freeman, Garrett Graham, Bill Hankins, Tom Smith, and Huai-Min Zhang. Improvements of the daily optimum interpolation sea surface temperature (doisst) version 2.1.Journal of Climate, 34(8):2923–2939, 2021

    2021

  47. [47]

    An extensible benchmark suite for learning to simulate physical systems

    Karl Otness, Arvi Gjoka, Joan Bruna, Daniele Panozzo, Benjamin Peherstorfer, Teseo Schneider, and Denis Zorin. An extensible benchmark suite for learning to simulate physical systems. arXiv preprint arXiv:2108.07799, 2021

    work page arXiv 2021

  48. [48]

    Scoring rules for continuous probability distributions

    James E Matheson and Robert L Winkler. Scoring rules for continuous probability distributions. Management science, 22(10):1087–1096, 1976

    1976

  49. [49]

    Error bounds for flow matching methods

    Joe Benton, George Deligiannidis, and Arnaud Doucet. Error bounds for flow matching methods. Transactions on Machine Learning Research, 2024

    2024

  50. [50]

    Imagenet: A large- scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large- scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  51. [51]

    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

  52. [52]

    Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers

    Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. InEuropean Conference on Computer Vision, pages 23–40. Springer, 2024. 12

    2024

  53. [53]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780–8794, 2021

    2021

  54. [54]

    High- resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High- resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 13 A Dataset Details In this section, we provide the formal governing equations and technical implemen...

    work page arXiv 2022