Pith. sign in

REVIEW 3 major objections 5 minor 32 references

Cosmo3DFlow claims that wavelet-space flow matching reconstructs cosmological initial conditions up to 50× faster than diffusion models while matching or improving reconstruction quality.

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

2026-08-03 02:37 UTC pith:OM224LA2

load-bearing objection Genuine speed engineering with a real runtime table, but the posterior-sampling claim is unsupported and the fidelity gain over plain flow matching is thinner than advertised. the 3 major comments →

arxiv 2602.10172 v2 pith:OM224LA2 submitted 2026-02-10 astro-ph.IM cs.AI

Cosmo3DFlow: Wavelet Flow Matching for Spatial-to-Spectral Compression in Reconstructing the Early Universe

Md. Khairul Islam , Zeyu Xia , Ryan Goudjil , Jialu Wang , Arya Farahi , Judy Fox This is my paper
classification astro-ph.IM cs.AI
keywords flow matchingdiscrete wavelet transformgenerative modelingcosmological initial conditionsinverse problemdensity field reconstructionposterior samplingsparsity
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.

This paper tries to establish that combining the 3D discrete wavelet transform with flow matching makes conditional generation of the universe's initial density field dramatically cheaper without sacrificing accuracy. The authors argue that the wavelet transform turns the spatial sparsity of cosmological voids into spectral sparsity: empty regions become near-zero high-frequency coefficients, so the generative model spends its computation on filaments and halos rather than emptiness. Working in wavelet space enables a stable ODE with 100 integration steps instead of 1000 and a five-fold lower per-step cost, yielding samples of 128^3 initial conditions in about five seconds on a single GPU. The reported result is 46–50× faster sampling than a diffusion baseline while improving point-wise reconstruction error, cross-correlation, and power-spectrum fidelity.

Core claim

Cosmo3DFlow demonstrates that the 3D discrete wavelet transform converts the 'void problem' of cosmological density fields—most voxels empty, yet each costs equal compute—into spectral sparsity: voids collapse to near-zero high-frequency coefficients while structure concentrates in the coarse LLL band. By running rectified-flow matching entirely in this 8-channel, half-resolution coefficient space and conditioning on the wavelet-transformed z=0 observation, the model learns a velocity field whose ODE can be integrated with only 100 Euler steps (versus 1000 for the diffusion baseline), with each step about 5× cheaper because the network operates on 64^3 rather than 128^3 voxels. On test volum

What carries the argument

The central mechanism is wavelet-space flow matching: apply the 3D Haar wavelet transform (pairwise averaging and differencing along each axis) to both the noise sample and the conditioning field, stack the eight subbands as channels, and train a 3D U-Net to regress the rectified-flow velocity—the difference between target and noise—in that coefficient space. Wavelet scale-specific conditioning and cross-scale skip connections inject multi-resolution information at each U-Net level, and a log-space power-spectrum loss regularizes the generated fields to have physically correct density statistics.

Load-bearing premise

The result stands or falls on whether the network's output is genuinely a sample from the conditional posterior distribution of initial conditions given the observed field, rather than a low-error reconstruction from a collapsed or narrow distribution.

What would settle it

Generate, say, 50 samples from the same z=0 field and measure the voxel-wise standard deviation and the power spectrum of the residuals; if the sample-to-sample dispersion is near zero, or far below the variance expected from the Gaussian prior on modes that the data cannot constrain, then the method is not sampling the posterior and the scientific claim collapses.

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

If this is right

  • At 128^3 resolution, a single initial-condition sample takes about 5 seconds instead of ~4 minutes, so studies that need thousands of posterior samples become practical.
  • The wavelet-space formulation reduces peak GPU memory by roughly 89% relative to the diffusion baseline at the same resolution, lowering hardware requirements.
  • Consistent gains across resolutions 32^3, 64^3, and 128^3 and across test sets with varying cosmological parameters and primordial non-Gaussianity suggest the method is not tuned to one dataset.
  • Wavelet-space velocity fields are stable enough for Euler integration with 100 steps, which directly enables the speedup; this appears to be a property of the representation, not just network engineering.
  • The framework suggests that computational cost scales with structure rather than volume, so the speedup should grow with resolution as voids dominate ever more of the grid.

Where Pith is reading between the lines

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

  • Beyond the paper's own claims: if the samples are genuinely draws from the posterior, then a per-sample cost of a few seconds makes it feasible to run Monte Carlo analyses over hundreds of realizations per observed volume, something that was previously out of reach.
  • The paper's evaluation does not yet test whether conditional samples are diverse; measuring the variance across samples conditioned on one z=0 field and comparing it to the expected posterior width would determine whether the method is doing posterior sampling or conditional-mean reconstruction.
  • The wavelet-sparsity mechanism should transfer to any inverse problem where the target field is spatially sparse but has dense small-scale features, such as turbulent flows or medical tomographic volumes; a direct test would be to apply the same architecture to a non-cosmological dataset.

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

3 major / 5 minor

Summary. The paper introduces Cosmo3DFlow, a conditional flow-matching generative model for reconstructing 3D initial conditions of cosmological density fields from evolved z=0 observations. The method operates in the 3D Haar wavelet domain: the noisy state and the conditioning observation are both transformed with DWT, a 3D U-Net predicts the velocity field in wavelet space, and inference is a deterministic Euler ODE followed by IDWT. Training combines a wavelet-space flow-matching loss with a log power-spectrum regularization term. Using Quijote N-body simulations at resolutions 32^3, 64^3, and 128^3, the paper reports roughly 46-50x faster sampling than a score-based diffusion baseline, with lower VRMSE and higher cross-correlation on three Quijote datasets. The contributions are framed as addressing the dimensionality and sparsity bottlenecks of cosmological inference through wavelet-based spatial-to-spectral compression.

Significance. If the speed and fidelity numbers are reproducible, this is a practically meaningful step for field-level cosmological inference: generating a 128^3 initial-condition sample in about 5 seconds on one A100, compared with minutes for diffusion baselines, would make posterior-style studies substantially more tractable. The wavelet-space formulation is well motivated by the void-dominated structure of cosmic density fields, and the paper uses public Quijote simulations, reports wall-clock and memory measurements, and gives enough architectural detail to attempt reproduction. However, the central interpretation of the method as sampling the posterior p(x|y) is not supported by any diversity or coverage metric; all reported quality metrics are pointwise or scale-wise aggregates. In addition, the power-spectrum R^2 metric is partly aligned with the training objective, and the quantitative comparison lacks error bars and a pixel-space flow-matching quality baseline at the maximum resolution. The result is therefore a promising but not yet fully validated contribution.

major comments (3)
  1. [§3.1, §4.1.2, Algorithm 2] The paper defines the task as sampling from the posterior p(x|y), but all reported metrics (VRMSE, cross-correlation, power-spectrum R^2, transfer function) evaluate only the fidelity of a single generated field against the truth. No experiment checks ensemble diversity, coverage, posterior consistency, or sensitivity to the initial noise draw. Because Algorithm 2 is a deterministic ODE from a single Gaussian draw, a model trained with an L2 flow-matching loss plus power-spectrum regularization can collapse toward a conditional mean or a narrow mode while still achieving excellent pointwise metrics. A concrete fix: fix a set of test observations y, draw multiple x0 for each y, and report the per-mode sample variance of the outputs alongside the truth variance, or compute a calibration/coverage diagnostic. Without such evidence, the phrase 'sampling in seconds' overstates what is demonstr
  2. [Eq. (10)-(11), Algorithm 1 line 12, Table 2] The headline quality claim 'Power Spectrum R^2-score' is partly circular. The training objective includes a log power-spectrum loss LPS that explicitly penalizes deviations of IDWT(M ⊙ v_t) from IDWT(M ⊙ (w1-w0)), so a high R^2 between the generated sample's power spectrum and the true power spectrum is expected from the regularization itself and is not an independent confirmation of fidelity. In addition, the mask M is never defined, and the loss is applied to the velocity field rather than to the final sample, so the relationship between P_pred = PowerSpec(IDWT(M ⊙ v_t)) and the power spectrum of the actual output sample is not established. The authors should define M, state where the power-spectrum R^2 is computed (final sample vs. training-time velocity), and report this metric for a model trained without the PS regularizer as an ablation.
  3. [Tables 2 and 4, Figure 4] The quantitative comparisons have gaps and internal inconsistencies that need to be resolved before the speed/fidelity claim is fully convincing. (i) Table 2 reports no error bars or repeated-seed statistics; only a single diffusion baseline is used. (ii) There is no pixel-space flow-matching quality baseline at 128^3; the ablation in Table 3 is at 32^3 only, so the claim that wavelet space 'matches or improves' quality for the high-resolution case is not directly tested against non-wavelet flow matching. (iii) The abstract and §5 say 'up to 50x' while Table 4 reports 46.8x at 128^3 and Figure 4 reports '~46x'; the abstract then says '10x fewer steps with 5x per-step speedup', but Table 4 implies 4.67x per-step speedup at 128^3. (iv) Figure 4 reports VRMSE 0.44 for Cosmo3DFlow at 128^3 while Table 2 reports 0.50 for the same resolution and dataset; this should be reconciled. (v) The diff
minor comments (5)
  1. [Abstract and §5] The abstract and conclusion state 'up to 50x faster sampling' while Table 4 reports 46.8x at the headline 128^3 resolution; please make the reported multiplier consistent across abstract, main text, and figures.
  2. [Table 2] The 'Transfer Function' column is labeled with an upward arrow, but the metric defined in §4.1.2 is the mean absolute deviation of T(k) from 1, for which lower is better. Please clarify the direction or change the label.
  3. [§3.3.1 and Eq. (6)] The normalization convention for the Haar DWT should be stated explicitly. Equation (6) divides only c_LLL by sqrt(8), but the detail coefficients are not normalized; the exact filter convention matters for the energy interpretation and for reproducibility.
  4. [Figure 6] The fiducial cosmology in the Figure 6 header (Omega_m=0.3, h=0.8, etc.) does not match the values stated in §4.1.1 (Omega_m=0.3175, h=0.6711). Please correct or explain.
  5. [Notation] The notation x for initial conditions and y for observations is opposite to the convention used in much of the cosmology inference literature (where y is usually the data and x the field to infer). Consider adding an explicit note or switching to a more standard convention to avoid confusion.

Circularity Check

1 steps flagged

One quality metric (power-spectrum R²) is partly a regularized training target; the central speed and VRMSE/correlation claims remain independent.

specific steps
  1. fitted input called prediction [§3.3.3 (Eq. 10–11), Algorithm 1 lines 12–15, §4.1.2, Table 2]
    "LPS = Σ_i (log P_pred(k_i) − log P_target(k_i))^2 ... Ltotal = Lflow + λPS LPS ... We keep the value at 0.01. ... Power Spectrum R2-score: ... we calculate the R2-score of the target 3D density field and the predicted samples. This shows how the model performs over the power spectrum."

    The headline metric 'Power Spectrum R²-score' reports how well the generated final density field's P(k) matches the target. The training objective includes a power-spectrum regularizer (Eq. 10) whose stated purpose is to enforce target power spectra, with λ_PS=0.01 in L_total. Although Algorithm 1 applies the regularizer to velocity fields rather than final fields, the paper motivates it as ensuring 'physical statistics', and the reported R² metric measures the same physical quantity on final fields. Therefore 'accurate power spectrum recovery' is partly a report that the regularized objective was optimized, not an independent test. This does not affect the runtime or VRMSE/cross-correlation evidence.

full rationale

No load-bearing self-citation chains are present: references to wavelet diffusion and flow matching are external prior work. The DWT is an invertible fixed transform, so the claimed 8× spatial compression is a representation choice, not a fitted result. The central speed comparison is grounded in measured runtimes and step counts on the same GPU, and VRMSE and cross-correlation are not direct training losses, providing independent fidelity evidence. The only noticeable reduction is the power-spectrum R² component of the reported quality: it is closely aligned with the power-spectrum regularization term in L_total, so that particular headline metric is partly self-reported training success rather than an independent validation. Separately, the paper frames the task as posterior sampling (§3.1) but evaluates only pointwise/aggregate fidelity (§4.1.2); this is a scientific validation gap, not a circularity, because the evaluation metrics are not defined in terms of the posterior-sampling objective. Score 4 reflects one partially circular quality metric while the central speed and pointwise-fidelity claims remain independent.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard flow-matching and DWT mathematics plus several domain assumptions about Quijote simulations and the sufficiency of power-spectrum regularization. The undefined mask M is the weakest ledger item: the power-spectrum loss and the reported P(k) improvement depend on it, but it is never specified.

free parameters (4)
  • Power-spectrum regularization weight λ_PS = 0.01
    Section 3.3.3: 'We keep the value at 0.01.' Chosen by hand; affects the physics-regularization loss and the reported P(k) metric.
  • Number of ODE steps K = 100 (Cosmo3DFlow), 1000 (diffusion baseline)
    Inference hyperparameters; the central 10×/50× speedup claim is defined by this ratio (§4.2, Table 4, Figure 4). Not fitted, but the speed claim is sensitive to it.
  • Wavelet level L = 1
    Algorithm 1 sets L←1; a single-level transform gives 8× spatial compression. Deeper levels or thresholding would change the sparsity/speed trade-off; chosen by hand.
  • Mask M in power-spectrum loss = undefined
    Algorithm 1 lines 12-13 multiply wavelet velocities by M before IDWT/PowerSpec; M is never defined in the text, Table 5, or appendix. The power-spectrum regularization term depends on this unspecified quantity.
axioms (5)
  • standard math Haar DWT is an orthonormal, invertible transform with perfect reconstruction; linear interpolation in wavelet space equals linear interpolation in pixel space.
    §3.3.1-3.3.2: w_t = t w1 + (1−t) w0 and IDWT recovery; because DWT is linear, the flow path is equivalent to pixel-space rectified flow.
  • domain assumption Initial conditions at z=127 are the correct target, and the Quijote simulation suite provides representative samples of the prior p(x).
    §4.1.1: training on Quijote LH/BSQ/fNL suites; the 'truth' is the N-body initial snapshot, not a physical observation.
  • domain assumption The z=0 logarithmic-density field carries sufficient information to condition on p(x|y); the inverse problem is learnable.
    §3.1: y = f(x) nonlinear forward model; the paper assumes a trained conditional generative model can approximate the posterior from these processed observations.
  • domain assumption Power spectrum is a sufficient statistic for Gaussian initial conditions, so power-spectrum regularization enforces the right physics.
    §3.3.3: 'Since the initial conditions are a Gaussian random field, the power spectrum serves as a sufficient statistic.' The loss is applied to velocities, not to final fields; this step is assumed to transfer.
  • ad hoc to paper There exists a well-defined mask M such that IDWT(M ⊙ v) power spectrum is a meaningful target for regularization.
    Algorithm 1 lines 12-13 use M without definition; the regularization and ablation results depend on this unspecified operator.

pith-pipeline@v1.3.0-alltime-deepseek · 15194 in / 15431 out tokens · 153765 ms · 2026-08-03T02:37:27.955666+00:00 · methodology

0 comments
read the original abstract

Reconstructing the early universe from the evolved present-day universe is a challenging and computationally demanding problem in modern astrophysics. We devise a novel generative framework, Cosmo3DFlow, designed to address dimensionality and sparsity, the critical bottlenecks inherent in current state-of-the-art methods for cosmological inference. By integrating 3D Discrete Wavelet Transform (DWT) with flow matching, we effectively represent high-dimensional cosmological structures. The Wavelet Transform addresses the ``void problem'' by translating spatial emptiness into spectral sparsity. It decouples high-frequency details from low-frequency structures, and wavelet-space velocity fields facilitate stable ordinary differential equation (ODE) solvers with large step sizes. Using large-scale cosmological $N$-body simulations at $128^3$ resolution, we achieve up to $46\times$ faster sampling than diffusion models. Our results enable initial conditions to be sampled in seconds, compared to minutes for previous methods.

Figures

Figures reproduced from arXiv: 2602.10172 by Arya Farahi, Jialu Wang, Judy Fox, Md. Khairul Islam, Ryan Goudjil, Zeyu Xia.

Figure 1
Figure 1. Figure 1: Schematic comparison of data representations for [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Cosmo3DFlow targets a high-dimensional inverse problem, while achieving 50 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of reconstruction quality. 2D slices through 3D density fields at [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Computational efficiency vs. reconstruction quality: [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The generation process for a test sample. Show [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The top row shows an example input, observa [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: 3D U-Net Velocity Network Architecture for res [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗

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

32 extracted references · 1 canonical work pages

  1. [1]

    J. M. Bardeen, J. R. Bond, N. Kaiser, and A. S. Szalay. 1986. The Statistics of Peaks of Gaussian Random Fields.The Astrophysical Journal304 (May 1986), 15. doi:10.1086/164143

    doi:10.1086/164143 1986

  2. [2]

    Atrideb Chatterjee and Francisco Villaescusa-Navarro. 2025. Cosmology from Point Clouds with Dark Matter Halos from the Quijote Simulations.The Astro- physical Journal985, 1 (May 2025), 132. doi:10.3847/1538-4357/adc99d

    work page doi:10.3847/1538-4357/adc99d 2025

  3. [3]

    2020.Modern Cosmology

    Scott Dodelson and Fabian Schmidt. 2020.Modern Cosmology. Academic Press, Cambridge, MA. doi:10.1016/C2017-0-01943-2

    doi:10.1016/c2017-0-01943-2 2020

  4. [4]

    Cat- tin

    Paul Friedrich, Julia Wolleb, Florentin Bieder, Alicia Durrer, and Philippe C. Cat- tin. 2025. WDM: 3D Wavelet Diffusion Models for High-Resolution Medical Image Synthesis. InDeep Generative Models, Anirban Mukhopadhyay, Ilkay Ok- suz, Sandy Engelhardt, Dorit Mehrof, and Yixuan Yuan (Eds.). Springer Nature Switzerland, Cham, 11–21. doi:10.1007/978-3-031-72744-3_2

    doi:10.1007/978-3-031-72744-3_2 2025

  5. [5]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising Diffusion Probabilistic Models. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., Vancouver, Canada, 6840–6851

    2020

  6. [6]

    2023.A Course in Cosmology: From Theory to Practice

    Dragan Huterer. 2023.A Course in Cosmology: From Theory to Practice. Cambridge University Press, Cambridge, UK

    2023

  7. [7]

    Jasche and G

    J. Jasche and G. Lavaux. 2019. Physical Bayesian Modelling of the Non-Linear Matter Distribution: New Insights into the Nearby Universe.Astronomy & Astrophysics625 (May 2019), A64. doi:10.1051/0004-6361/201833710

    doi:10.1051/0004-6361/201833710 2019

  8. [8]

    Jens Jasche and Benjamin D. Wandelt. 2013. Bayesian Physical Reconstruction of Initial Conditions from Large-Scale Structure Surveys.Monthly Notices of the Royal Astronomical Society432, 2 (June 2013), 894–913. doi:10.1093/mnras/stt449

    doi:10.1093/mnras/stt449 2013

  9. [9]

    Sidharth Kannan, Tian Qiu, Carolina Cuesta-Lazaro, and Haewon Jeong. 2025. CosmoFlow: Scale-Aware Representation Learning for Cosmology with Flow Matching. arXiv:2507.11842 [astro-ph]

    Pith/arXiv arXiv 2025

  10. [10]

    Ronan Legin, Matthew Ho, Pablo Lemos, Laurence Perreault-Levasseur, Shirley Ho, Yashar Hezaveh, and Benjamin Wandelt. 2024. Posterior Sampling of the Initial Conditions of the Universe from Non-Linear Large Scale Structures Using Score-Based Generative Models.Monthly Notices of the Royal Astronomical Society: Letters527, 1 (Jan. 2024), L173–L178. doi:10.1...

    doi:10.1093/mnrasl/slad152 2024

  11. [11]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. 2023. Flow Matching for Generative Modeling. arXiv:2210.02747 [cs]

    Pith/arXiv arXiv 2023

  12. [12]

    Xingchao Liu, Chengyue Gong, and Qiang Liu. 2022. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. arXiv:2209.03003 [cs]

    Pith/arXiv arXiv 2022

  13. [13]

    Mebratu and W

    M. Mebratu and W. L. K. Wu. 2025. Wavelet Flow For Extragalactic Foreground Simulations. arXiv:2505.21220 [astro-ph]

    Pith/arXiv arXiv 2025

  14. [14]

    Libeskind, Yehuda Hoffman, Robert A

    Ofer Metuki, Noam I. Libeskind, Yehuda Hoffman, Robert A. Crain, and Tom Theuns. 2015. Galaxy Properties and the Cosmic Web in Simulations.Monthly Notices of the Royal Astronomical Society446, 2 (Jan. 2015), 1458–1468. doi:10. 1093/mnras/stu2166

    2015

  15. [15]

    Satvik Mishra, Roberto Trotta, and Matteo Viel. 2026. Cosmo-FOLD: Fast Gen- eration and Upscaling of Field-Level Cosmological Maps with Overlap Latent Diffusion. arXiv:2601.14377 [astro-ph] doi:10.48550/arXiv.2601.14377

    doi:10.48550/arxiv.2601.14377 2026

  16. [16]

    C. Modi, F. Lanusse, and U. Seljak. 2021. FlowPM: Distributed TensorFlow Imple- mentation of the FastPM Cosmological N-body Solver.Astronomy and Computing 37 (Oct. 2021), 100505. doi:10.1016/j.ascom.2021.100505

    arXiv 2021

  17. [17]

    Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B

    Ruben Ohana, Michael McCabe, Lucas Meyer, Rudy Morel, Fruzsina J. Agocs, Miguel Beneitez, Marsha Berger, Blakesley Burkhart, Stuart B. Dalziel, Drum- mond B. Fielding, Daniel Fortunato, Jared A. Goldberg, Keiya Hirashima, Yan-Fei Jiang, Rich R. Kerswell, Suryanarayana Maddu, Jonah Miller, Payel Mukhopad- hyay, Stefan S. Nixon, Jeff Shen, Romain Watteaux, ...

    2024

  18. [18]

    Hao Phung, Quan Dao, and Anh Tran. 2023. Wavelet Diffusion Models Are Fast and Scalable Image Generators. In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, Vancouver, BC, Canada, 10199–10208. doi:10.1109/CVPR52729.2023.00983

    arXiv 2023

  19. [19]

    Julieth K Riveros, Paola A Saavedra, Héctor J Hortúa, Jorge Enrique García- Farieta, and Ivan Olier. 2025. Conditional Diffusion-Flow models for generating 3D cosmic density fields: applications to f (R) cosmologies.Machine Learning: Science and Technology6, 3 (2025), 035031

    2025

  20. [20]

    François Rozet, Ruben Ohana, Michael McCabe, Gilles Louppe, François Lanusse, and Shirley Ho. 2025. Lost in Latent Space: An Empirical Study of Latent Diffusion Models for Physics Emulation. arXiv:2507.02608 [cs]

    arXiv 2025

  21. [21]

    Toomey, and Sergei Gleyzer

    Hamees Sayed, Pranath Reddy, Michael W. Toomey, and Sergei Gleyzer

  22. [22]

    Christopher J Shallue and Daniel J Eisenstein. 2023. Reconstructing Cosmological Initial Conditions from Late-Time Structure with Convolutional Neural Networks. Monthly Notices of the Royal Astronomical Society520, 4 (April 2023), 6256–6267. doi:10.1093/mnras/stad528

    doi:10.1093/mnras/stad528 2023

  23. [23]

    Luigi Sigillo, Shengfeng He, and Danilo Comminiello. 2025. Latent Wavelet Diffusion: Enabling 4K Image Synthesis for Free. arXiv:2506.00433 [cs]

    Pith/arXiv arXiv 2025

  24. [24]

    Yang Song and Stefano Ermon. 2020. Improved Techniques for Training Score- Based Generative Models. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., Vancouver, Canada, 12438–12448

    2020

  25. [25]

    Matthew Tancik, Pratul Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan Barron, and Ren Ng

  26. [26]

    Georgios Valogiannis, Francisco Villaescusa-Navarro, and Marco Baldi. 2024. Towards unveiling the large-scale nature of gravity with the wavelet scattering transform.Journal of Cosmology and Astroparticle Physics2024, 11 (2024), 061

    2024

  27. [27]

    Francisco Villaescusa-Navarro. 2018. Pylians: Python libraries for the analysis of numerical simulations. Astrophysics Source Code Library, record ascl:1811.008, ascl:1811.008 pages. ascl:1811.008

    2018

  28. [28]

    Kreisch, Andrina Nicola, Justin Alsing, Roman Scoccimarro, Licia Verde, Matteo Viel, Shirley Ho, Stephane Mallat, Benjamin Wandelt, and David N

    Francisco Villaescusa-Navarro, ChangHoon Hahn, Elena Massara, Arka Banerjee, Ana Maria Delgado, Doogesh Kodi Ramanah, Tom Charnock, Elena Giusarma, Yin Li, Erwan Allys, Antoine Brochard, Cora Uhlemann, Chi-Ting Chiang, Siyu He, Alice Pisani, Andrej Obuljen, Yu Feng, Emanuele Castorina, Gabriella Contardo, Christina D. Kreisch, Andrina Nicola, Justin Alsin...

    2020

  29. [29]

    Simon D. M. White and Carlos S. Frenk. 1991. Galaxy Formation through Hier- archical Clustering.The Astrophysical Journal379 (Sept. 1991), 52. doi:10.1086/ 170483

    1991

  30. [30]

    Maksim Zhdanov, Max Welling, and Jan-Willem van de Meent. 2025. Er- win: A Tree-based Hierarchical Transformer for Large-scale Physical Systems. arXiv:2502.17019 [cs] A Appendix Table 5 summarizes the notation and key definitions used through- out this paper. 3D U-Net Velocity Network ODE Sampling Euler, K steps 32ch 32³ 64ch 16³ 128ch 8³ 64ch 16³ 32ch 32...

    Pith/arXiv arXiv 2025

  31. [2020]

    InAdvances in Neural Information Processing Systems, Vol

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains. InAdvances in Neural Information Processing Systems, Vol. 33. Curran Associates, Inc., Vancouver, BC, Canada, 7537–7547

  32. [2025]

    arXiv:2510.07878 [astro-ph]

    FlowLensing: Simulating Gravitational Lensing with Flow Matching. arXiv:2510.07878 [astro-ph]

    arXiv