Pith. sign in

REVIEW 2 major objections 4 minor 48 references

A diffusion model can generate galaxy catalogues with correct clustering statistics directly from a dark matter density field, without ever identifying haloes.

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

2026-08-01 11:37 UTC pith:DBJXK4WX

load-bearing objection A practical, honest step forward in diffusion-based mock generation; the independent-patch assumption is the load-bearing caveat and higher-order statistics are untested. the 2 major comments →

arxiv 2607.19808 v1 pith:DBJXK4WX submitted 2026-07-22 astro-ph.GA astro-ph.CO

From Dark Matter to Galaxies: Halo-Free Mock Generation via Conditional Point-Cloud Diffusion

Kana Moriwaki , Ken Osato , Naoki Yoshida This is my paper
classification astro-ph.GA astro-ph.CO
keywords galaxy mocksdiffusion modelspoint cloud generationdark matter density fieldgalaxy clusteringpower spectrumstar formation ratecosmological simulations
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 authors aim to show that a conditional diffusion model can generate realistic galaxy mocks directly from a dark matter density field, bypassing the traditional step of halo finding. They train the model on a hydrodynamical simulation and demonstrate that the generated galaxies reproduce the one- and two-dimensional distributions of star formation rate and stellar mass, as well as the galaxy auto-power spectrum and the galaxy–dark matter cross-power spectrum. The method also produces galaxies hosted by structures too small to be resolved in the input density field, by statistically marginalizing over unresolved small-scale structure. Because generation runs in about ten seconds for a (151 Mpc)^3 volume on a single GPU, the approach could enable cheap, large ensembles of mocks for upcoming surveys.

Core claim

The paper's central claim is that a point-cloud diffusion model conditioned on a low-resolution dark matter density field can act as a direct, halo-free mapping from matter to galaxies. The model inputs a 3D density grid with ~1.2 Mpc voxels, downsampled to mimic a very low-resolution dark-matter-only simulation, and outputs galaxy points with positions, star formation rate, and stellar mass. Trained on a hydrodynamical simulation, the generated catalogs match the reference simulation's galaxy properties, clustering amplitude, and cross-correlation with the dark matter field over a wide range of scales. The authors further show that the model generates galaxies in halos below the nominal res

What carries the argument

The central object is a conditional point-cloud diffusion model built on the Elucidated Diffusion Model (EDM) framework: a transformer-based denoising network that takes noisy galaxy point sets and a voxelized dark matter density cube as input, and learns to denoise the points into realistic galaxy configurations with physical attributes. Self-attention among galaxy tokens captures galaxy–galaxy correlations, while cross-attention to the density cube injects the dark matter environment. The independent-patch generation with tiling is the mechanism that makes large volumes tractable, and the stochastic sampling step is what enables the model to marginalize over unresolved small-scale structur

Load-bearing premise

The method assumes that galaxy formation is governed primarily by the local dark matter density within patches of about 19 Mpc, so that galaxies in one patch are statistically independent of everything outside it; if long-range environmental effects or correlations beyond a few megaparsecs matter, the generated mocks will miss them.

What would settle it

Measure the galaxy bispectrum (or three-point correlation function) of the generated catalogue against the reference simulation over the full test volume; a significant deviation on scales near or above the patch size would falsify the independent-patch assumption and reveal missing large-scale mode coupling.

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

If this is right

  • If the approach holds up, mock galaxy catalogues for redshift surveys and line-intensity mapping can be generated directly from cheap, low-resolution dark-matter-only simulations, removing the need for halo finders and subhalo abundance matching.
  • The ability to generate galaxy populations below the resolution limit means that faint or low-mass galaxies, which dominate future intensity-mapping signals, can be modelled from density fields that do not resolve their host halos.
  • Because each patch is generated independently, the method is naturally parallelizable and can scale to arbitrarily large survey volumes by tiling.
  • The point-cloud output is a continuous representation, allowing the catalogues to be trimmed or reshaped to survey geometries and selection functions without a fixed grid.
  • Extending the per-galaxy feature vector (e.g., adding velocities, gas masses, metallicities) is computationally cheap in this framework, making richer mock observables feasible at little extra cost.

Where Pith is reading between the lines

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

  • The paper's patch-independence assumption (that galaxy formation is local and correlations die out within about 20 Mpc) is only validated for two-point statistics; if environmental effects like assembly bias or tidal alignment act on larger scales, they would show up first in the bispectrum or marked correlation functions, so those should be tested before using the mocks for precision cosmology.
  • The observed ability to marginalize over unresolved structure suggests the model could be trained on even lower-resolution inputs, such as fast particle-mesh simulations, and still produce believable small-scale galaxy populations — a testable extension that would further lower the cost of mock production.
  • Since the model learns a conditional distribution rather than a deterministic map, the generated realizations naturally encode the stochastic scatter of galaxy formation, which could be exploited for simulation-based inference where field-level likelihoods are needed.
  • The patch size is a hyperparameter; training the same model on larger conditioning volumes (say 50–100 Mpc) would directly test whether large-scale modes beyond the current patch are needed for high-order statistics, and would set a practical limit on the method's accuracy.

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

2 major / 4 minor

Summary. The paper presents a conditional diffusion model (EDM) that generates galaxy point clouds — positions, star formation rate, and stellar mass — directly from a three-dimensional dark-matter density field, without any halo or subhalo identification. The model is trained on IllustrisTNG300 at z=2, using a heavily downsampled dark-matter density field as input (1.2 Mpc voxels, 16^3 voxels per 18.9 Mpc patch) and TNG300-1 galaxies with SFR > 1 M_sun/yr as targets. Generation is performed patch-by-patch and the patches are tiled to cover a held-out (151.3 Mpc)^3 test volume. The paper reports that the generated catalogues reproduce the one- and two-dimensional distributions of SFR and stellar mass, the galaxy auto-power spectrum, and the galaxy–DM cross-power spectrum, and that the model statistically populates structures below the nominal resolution limit of the input density field. An optimized sampling schedule yields a ~10 second generation time for the full test volume. An autoregressive patch-generation variant and sampling-schedule tuning are presented in appendices.

Significance. If the claims hold, this is a useful proof-of-concept for fast, halo-free mock generation that could be relevant for survey ensembles and simulation-based inference. The clean train/test split, the use of multiple realizations for property distributions, the explicit comparison to a halo-mass-threshold baseline, and the careful tuning of the sampling schedule are strengths. The paper is appropriately cautious in stating that two-point agreement does not guarantee higher-order fidelity. However, the quantitative support for the central clustering claim is weakened by the absence of error bars on the power-spectrum ratios, and the stated survey-mock motivation goes beyond what is actually validated.

major comments (2)
  1. [§3.2, Fig. 4] The auto- and cross-power spectra are displayed as single ratio curves with no uncertainty. The abstract and Section 3.2 claim that the generated catalogues 'accurately reproduce' these spectra, but with a single realization the reader cannot assess the scatter, especially on large scales where sample variance is significant. Since 100 independent realizations are already used for Fig. 3, the authors should show the mean and standard deviation (or an error band) of P_gen and P_X,gen over these realizations, and state how many realizations are used for the power-spectrum estimate. This is needed to support the central clustering claim.
  2. [§2.1, §5, Abstract] Patch-wise independent generation is validated only for two-point statistics; the paper explicitly acknowledges in §3.2 that higher-order correlations and mode-mode coupling are not guaranteed. However, the abstract and Introduction motivate the method as a 'practical engine for producing large mock ensembles' for surveys, which rely on accurate covariance matrices and non-Gaussian statistics. This is a mismatch between the validated scope and the stated applications. Either add a higher-order test (e.g., reduced bispectrum or another non-Gaussian statistic) or explicitly scope the claims to two-point statistics and temper the abstract/motivation accordingly.
minor comments (4)
  1. [§2.1] 'containing ∼2903 DM particles' appears to be a typo; presumably 290^3 or another power-of-ten count is intended. Please correct the missing superscript.
  2. [§3.2] The sentence 'The black dashed line show that of TNG galaxies...' should be 'shows'; also note that the dashed curve uses a 10^12 M_sun/h halo-mass threshold, which corresponds to ~20 particles in the downsampled input, while §2.1 states that reliable haloes require ~100 particles (M_h ≳ 4.5×10^12). Clarifying this would prevent over-interpretation of the comparison.
  3. [Fig. 4 and Fig. A2] The x-axis label 'k [Mpc 1]' should be 'k [Mpc^{-1}]'.
  4. [Eq. (1)] The normalization expression is ambiguous: '2 log s − a b − a − 1' should be written with clear parentheses, e.g., 2(log s − a)/(b − a) − 1.

Circularity Check

0 steps flagged

No significant circularity: the method is trained on TNG patches and validated against a held-out TNG volume with independently selected hyperparameters.

full rationale

The paper's central derivation is self-contained and not circular. The conditional diffusion model is trained on patches drawn from IllustrisTNG and evaluated on a reserved (151.3 Mpc)^3 test volume that was not used in training or validation; the reported SFR/stellar-mass distributions, auto-power spectrum, and cross-power spectrum are measured from generated samples and compared directly to TNG data. No parameter is fitted to these test statistics: the sampling schedule parameter alpha=0.6 is chosen on validation data via the Wasserstein distance in Fig. B2, and the number of sampling steps is also a validation decision, not a fit to the final power spectra. The comparison against the resolved-halo baseline (black dashed line in Fig. 3) is a diagnostic, not an input. The self-citations to Moriwaki et al. (2026) and Osato & Okumura (2023) are used only as context or comparison (related work and a general statement about hydrodynamical simulation mocks) and do not supply any load-bearing premise for the results shown. The paper's own caveat that two-point agreement does not guarantee higher-order correlations (Section 3.2) is an acknowledged limitation rather than a circular step. Thus there is no identifiable reduction of any prediction to its own inputs, and no circularity score above 0 is warranted.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper relies on TNG as ground truth and a downsampled DM field as input; these are domain assumptions. No new physical entities are introduced. The free parameters are preprocessing choices and hyperparameters.

free parameters (5)
  • SFR normalization bounds (a,b) = (0,3)
    Hand-chosen to map the typical SFR dynamic range to approximately [-1,1] (Eq. 1).
  • Stellar mass normalization bounds (a,b) = (7,12)
    Hand-chosen to map the typical stellar mass dynamic range to approximately [-1,1] (Eq. 1).
  • DM density normalization factor = log10(rho)/5
    Hand-chosen to scale the density values to approximately [0,1] (Section 2.1).
  • Sampling schedule parameter alpha = 0.6
    Chosen on validation data (Wasserstein-1 distance) to balance quality and cost; not fitted to the final test result (Appendix B).
  • Patch size = 18.9 Mpc
    Chosen as a trade-off between computational cost and capturing local physics; underlies the locality assumption (Section 2.1).
axioms (4)
  • domain assumption The TNG simulation is a reliable representation of galaxy formation.
    Used as ground truth for training and validation; the model's realism is only as good as TNG's subgrid models.
  • domain assumption The downsampled TNG300-3-Dark density field is a valid proxy for a low-resolution DM-only simulation.
    Section 2.1; if this proxy differs from true low-resolution simulations (force softening, initial conditions), the model may not generalize.
  • domain assumption Galaxy formation is primarily governed by local physical conditions within a patch of ~19 Mpc.
    Section 2.1; central to the patch-independent generation strategy and scalability.
  • domain assumption Stochasticity in star formation justifies comparing statistical distributions rather than individual objects.
    Section 3.1; the model samples from a conditional distribution and is not expected to match TNG galaxy-by-galaxy.

pith-pipeline@v1.3.0-alltime-deepseek · 12859 in / 11465 out tokens · 116992 ms · 2026-08-01T11:37:50.796325+00:00 · methodology

0 comments
read the original abstract

We present a diffusion-based generative model for constructing realistic galaxy catalogues from dark matter density fields. The model takes a three-dimensional dark matter density field as input and generates galaxies directly as a point cloud with positions and physical properties, including star formation rate (SFR), without identifying dark matter haloes or subhaloes. We train the model on galaxy catalogues from the IllustrisTNG hydrodynamical simulation. The generated catalogues reproduce the spatial correspondence between galaxies and the underlying dark matter field, preferentially populating dense regions and filamentary structures. They also accurately reproduce the one- and two-dimensional distributions of SFR and stellar mass, the galaxy auto-power spectrum, and the galaxy--dark matter cross-power spectrum. The model generates galaxies associated with structures below the nominal resolution of the input density field by marginalising over unresolved small-scale structure rather than relying on a resolved halo catalogue. With an optimised diffusion sampling schedule, it generates a catalogue with SFR $> 1 ~\rm M_\odot/yr$ over a $(151.3 ~\rm Mpc)^3$ volume in approximately 10 seconds on a single GPU. Our model therefore provides a practical engine for producing large mock ensembles for upcoming galaxy redshift surveys and line-intensity mapping experiments, and offers a path toward simulation-based inference that bypasses halo finding and directly connects field-level dark matter statistics to observable galaxy populations.

Figures

Figures reproduced from arXiv: 2607.19808 by Kana Moriwaki, Ken Osato, Naoki Yoshida.

Figure 1
Figure 1. Figure 1: True and generated point distributions in patches randomly extracted from the test region. Each patch has a side length of 18.9 Mpc. 3 RESULTS 3.1 Visual Inspection [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: DM density (left), true TNG galaxies (middle), and generated galaxies (right) in the test region of a volume of (151.3 Mpc) 3 10 0 10 1 10 2 10 3 SFR [M /yr] 10 5 10 4 10 3 d n/dlo g S F R [M p c 3 d e x 1 ] TNG TNG (Mh > 10 12 M /h) Generated 10 10 10 11 10 12 Mstar [M ] 10 5 10 4 10 3 d n/dlo g Mstar [M p c 3 d e x 1 ] TNG TNG (Mh > 10 12 M /h) Generated 10 7 10 8 10 9 10 10 10 11 10 12 Mstar [M ] 10 0 1… view at source ↗
Figure 3
Figure 3. Figure 3: Histograms of SFRs (left) and stellar masses (middle) of TNG (black) and generated galaxies (red). For the generated ones, the shaded regions show the variance of 100 different realizations. For comparison, the black dashed line show that of TNG galaxies with halo mass greater than 1012 M⊙/ℎ. Right panel shows the two-dimensional histogram. ies are correctly correlated with the underlying matter distributi… view at source ↗
Figure 5
Figure 5. Figure 5: Cross-correlation coefficient between the TNG and generated of galaxy distributions. For both TNG and generated catalog, galaxies with SFR > 1 M⊙/yr are used. et al. 2026), where modelling multiple correlated properties may require representing them as additional tokens and therefore increas￾ing the effective sequence length. Since the computational cost of transformer-based models is typically more sensit… view at source ↗
Figure 4
Figure 4. Figure 4: Top: Power spectrum of galaxy number density map. For both TNG and generated catalog, galaxies with SFR > 1 M⊙/yr are used. The vertical dotted line indicates the patch-size scale 𝑘patch = 2𝜋/𝑙patch. Bottom: Cross-power spectrum between galaxy number density map and DM density map. The upper panels show the ratios of the predicted power spectra to the corresponding TNG results. catalogues across different … 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

48 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    Aguirre J., STARFIRE Collaboration 2018, in American Astronomical Society Meeting Abstracts \#231. p. 328.04

    2018

  2. [2]

    E., Cohan A., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2004.05150 , https://ui.adsabs.harvard.edu/abs/2020arXiv200405150B p

    Beltagy I., Peters M. E., Cohan A., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2004.05150 , https://ui.adsabs.harvard.edu/abs/2020arXiv200405150B p. arXiv:2004.05150

    Pith/arXiv arXiv doi:10.48550/arxiv.2004.05150 2020

  3. [3]

    CCAT-Prime Collaboration et al., 2023, @doi [ ] 10.3847/1538-4365/ac9838 , https://ui.adsabs.harvard.edu/abs/2023ApJS..264....7C 264, 7

    doi:10.3847/1538-4365/ac9838 2023

  4. [4]

    CONCERTO Collaboration et al., 2020, @doi [ ] 10.1051/0004-6361/202038456 , https://ui.adsabs.harvard.edu/abs/2020A&A...642A..60C 642, A60

    doi:10.1051/0004-6361/202038456 2020

  5. [5]

    T., et al., 2014, in Holland W

    Crites A. T., et al., 2014, in Holland W. S., Zmuidzinas J., eds, Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series Vol. 9153, Millimeter, Submillimeter, and Far-Infrared Detectors and Instrumentation for Astronomy VII. p. 91531W, @doi 10.1117/12.2057207

    doi:10.1117/12.2057207 2014

  6. [6]

    Cuesta-Lazaro C., Mishra-Sharma S., 2024, @doi [ ] 10.1103/PhysRevD.109.123531 , https://ui.adsabs.harvard.edu/abs/2024PhRvD.109l3531C 109, 123531

    work page doi:10.1103/physrevd.109.123531 2024

  7. [7]

    arXiv:1611.00036

    DESI Collaboration et al., 2016, @doi [arXiv e-prints] 10.48550/arXiv.1611.00036 , https://ui.adsabs.harvard.edu/abs/2016arXiv161100036D p. arXiv:1611.00036

    Pith/arXiv arXiv doi:10.48550/arxiv.1611.00036 2016

  8. [8]

    arXiv:2010.11929

    Dosovitskiy A., et al., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2010.11929 , https://ui.adsabs.harvard.edu/abs/2020arXiv201011929D p. arXiv:2010.11929

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2020

  9. [9]

    Euclid Collaboration et al., 2022, @doi [ ] 10.1051/0004-6361/202141938 , https://ui.adsabs.harvard.edu/abs/2022A&A...662A.112E 662, A112

    doi:10.1051/0004-6361/202141938 2022

  10. [10]

    Euclid Collaboration et al., 2025, @doi [ ] 10.1051/0004-6361/202450810 , https://ui.adsabs.harvard.edu/abs/2025A&A...697A...1E 697, A1

    doi:10.1051/0004-6361/202450810 2025

  11. [11]

    arXiv:2006.11239

    Ho J., Jain A., Abbeel P., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2006.11239 , https://ui.adsabs.harvard.edu/abs/2020arXiv200611239H p. arXiv:2006.11239

    Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  12. [12]

    K., Cranmer M., Melchior P., Ho S., Somerville R

    Jespersen C. K., Cranmer M., Melchior P., Ho S., Somerville R. S., Gabrielpillai A., 2022, @doi [ ] 10.3847/1538-4357/ac9b18 , https://ui.adsabs.harvard.edu/abs/2022ApJ...941....7J 941, 7

    doi:10.3847/1538-4357/ac9b18 2022

  13. [13]

    Jo Y., Kim J.-h., 2019, @doi [ ] 10.1093/mnras/stz2304 , https://ui.adsabs.harvard.edu/abs/2019MNRAS.489.3565J 489, 3565

    doi:10.1093/mnras/stz2304 2019

  14. [14]

    M., Turk M

    Kamdar H. M., Turk M. J., Brunner R. J., 2016, @doi [ ] 10.1093/mnras/stv2981 , https://ui.adsabs.harvard.edu/abs/2016MNRAS.457.1162K 457, 1162

    doi:10.1093/mnras/stv2981 2016

  15. [15]

    arXiv:2206.00364

    Karras T., Aittala M., Aila T., Laine S., 2022, @doi [arXiv e-prints] 10.48550/arXiv.2206.00364 , https://ui.adsabs.harvard.edu/abs/2022arXiv220600364K p. arXiv:2206.00364

    Pith/arXiv arXiv doi:10.48550/arxiv.2206.00364 2022

  16. [16]

    P., Ba J., 2014, @doi [arXiv e-prints] 10.48550/arXiv.1412.6980 , https://ui.adsabs.harvard.edu/abs/2014arXiv1412.6980K p

    Kingma D. P., Ba J., 2014, @doi [arXiv e-prints] 10.48550/arXiv.1412.6980 , https://ui.adsabs.harvard.edu/abs/2014arXiv1412.6980K p. arXiv:1412.6980

    Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2014

  17. [17]

    arXiv:0912.0201

    LSST Science Collaboration et al., 2009, @doi [arXiv e-prints] 10.48550/arXiv.0912.0201 , https://ui.adsabs.harvard.edu/abs/2009arXiv0912.0201L p. arXiv:0912.0201

    Pith/arXiv arXiv doi:10.48550/arxiv.0912.0201 2009

  18. [18]

    arXiv:2211.09958

    Li Y., et al., 2022, @doi [arXiv e-prints] 10.48550/arXiv.2211.09958 , https://ui.adsabs.harvard.edu/abs/2022arXiv221109958L p. arXiv:2211.09958

    Pith/arXiv arXiv doi:10.48550/arxiv.2211.09958 2022

  19. [19]

    arXiv:2103.14030

    Liu Z., Lin Y., Cao Y., Hu H., Wei Y., Zhang Z., Lin S., Guo B., 2021, @doi [arXiv e-prints] 10.48550/arXiv.2103.14030 , https://ui.adsabs.harvard.edu/abs/2021arXiv210314030L p. arXiv:2103.14030

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.14030 2021

  20. [20]

    C., et al., 2023, in Machine Learning for Astrophysics

    Lovell C. C., et al., 2023, in Machine Learning for Astrophysics. p. 21 ( @eprint arXiv 2307.06967 ), @doi 10.48550/arXiv.2307.06967

    Pith/arXiv arXiv doi:10.48550/arxiv.2307.06967 2023

  21. [21]

    Marinacci F., et al., 2018, @doi [ ] 10.1093/mnras/sty2206 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.480.5113M 480, 5113

    doi:10.1093/mnras/sty2206 2018

  22. [22]

    arXiv:2601.14377

    Mishra S., Trotta R., Viel M., 2026a, @doi [arXiv e-prints] 10.48550/arXiv.2601.14377 , https://ui.adsabs.harvard.edu/abs/2026arXiv260114377M p. arXiv:2601.14377

    doi:10.48550/arxiv.2601.14377

  23. [23]

    Mishra S., Trotta R., Viel M., 2026b, @doi [ ] 10.1093/mnras/staf2071 , https://ui.adsabs.harvard.edu/abs/2026MNRAS.545f2071M 545, staf2071

    doi:10.1093/mnras/staf2071

  24. [24]

    L., Osato K., Yoshida N., 2026, @doi [ ] 10.1093/mnras/staf2124 , https://ui.adsabs.harvard.edu/abs/2026MNRAS.545f2124M 545, staf2124

    Moriwaki K., Jun R. L., Osato K., Yoshida N., 2026, @doi [ ] 10.1093/mnras/staf2124 , https://ui.adsabs.harvard.edu/abs/2026MNRAS.545f2124M 545, staf2124

    doi:10.1093/mnras/staf2124 2026

  25. [25]

    P., et al., 2018, @doi [ ] 10.1093/mnras/sty618 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.477.1206N 477, 1206

    Naiman J. P., et al., 2018, @doi [ ] 10.1093/mnras/sty618 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.477.1206N 477, 1206

    doi:10.1093/mnras/sty618 2018

  26. [26]

    Nelson D., et al., 2018, @doi [ ] 10.1093/mnras/stx3040 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.475..624N 475, 624

    doi:10.1093/mnras/stx3040 2018

  27. [27]

    Nelson D., et al., 2019, @doi [Computational Astrophysics and Cosmology] 10.1186/s40668-019-0028-x , https://ui.adsabs.harvard.edu/abs/2019ComAC...6....2N 6, 2

    doi:10.1186/s40668-019-0028-x 2019

  28. [28]

    arXiv:2409.02980

    Nguyen T., et al., 2024, @doi [arXiv e-prints] 10.48550/arXiv.2409.02980 , https://ui.adsabs.harvard.edu/abs/2024arXiv240902980N p. arXiv:2409.02980

    Pith/arXiv arXiv doi:10.48550/arxiv.2409.02980 2024

  29. [29]

    Osato K., Okumura T., 2023, @doi [ ] 10.1093/mnras/stac3582 , https://ui.adsabs.harvard.edu/abs/2023MNRAS.519.1771O 519, 1771

    doi:10.1093/mnras/stac3582 2023

  30. [30]

    Oxholm T., et al., 2020, in American Astronomical Society Meeting Abstracts \#236. p. 244.02

    2020

  31. [31]

    C., Modi C., Wandelt B

    Pandey S., Lovell C. C., Modi C., Wandelt B. D., 2025, @doi [arXiv e-prints] 10.48550/arXiv.2511.08438 , https://ui.adsabs.harvard.edu/abs/2025arXiv251108438P p. arXiv:2511.08438

    work page doi:10.48550/arxiv.2511.08438 2025

  32. [32]

    arXiv:1912.01703

    Paszke A., et al., 2019, @doi [arXiv e-prints] 10.48550/arXiv.1912.01703 , https://ui.adsabs.harvard.edu/abs/2019arXiv191201703P p. arXiv:1912.01703

    Pith/arXiv arXiv doi:10.48550/arxiv.1912.01703 2019

  33. [33]

    arXiv:2212.09748

    Peebles W., Xie S., 2022, @doi [arXiv e-prints] 10.48550/arXiv.2212.09748 , https://ui.adsabs.harvard.edu/abs/2022arXiv221209748P p. arXiv:2212.09748

    Pith/arXiv arXiv doi:10.48550/arxiv.2212.09748 2022

  34. [34]

    Pillepich A., et al., 2018a, @doi [ ] 10.1093/mnras/stx2656 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.473.4077P 473, 4077

    doi:10.1093/mnras/stx2656

  35. [35]

    Pillepich A., et al., 2018b, @doi [ ] 10.1093/mnras/stx3112 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.475..648P 475, 648

    doi:10.1093/mnras/stx3112

  36. [36]

    arXiv:2112.10752

    Rombach R., Blattmann A., Lorenz D., Esser P., Ommer B., 2021, @doi [arXiv e-prints] 10.48550/arXiv.2112.10752 , https://ui.adsabs.harvard.edu/abs/2021arXiv211210752R p. arXiv:2112.10752

    Pith/arXiv arXiv doi:10.48550/arxiv.2112.10752 2021

  37. [37]

    Rouhiainen A., M \"u nchmeyer M., Shiu G., Gira M., Lee K., 2024, @doi [ ] 10.1103/PhysRevD.109.123536 , https://ui.adsabs.harvard.edu/abs/2024PhRvD.109l3536R 109, 123536

    doi:10.1103/physrevd.109.123536 2024

  38. [38]

    Probabilistic Galaxy Field Generation with Diffusion Models

    Sether T., Giusarma E., Reyes-Hurtado M., 2024, @doi [arXiv e-prints] 10.48550/arXiv.2412.05131 , https://ui.adsabs.harvard.edu/abs/2024arXiv241205131S p. arXiv:2412.05131

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2412.05131 2024

  39. [39]

    Shirasaki M., Ikeda S., 2024, @doi [The Open Journal of Astrophysics] 10.33232/001c.118104 , https://ui.adsabs.harvard.edu/abs/2024OJAp....7E..42S 7, 42

    doi:10.33232/001c.118104 2024

  40. [40]

    P., Kumar A., Ermon S., Poole B., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2011.13456 , https://ui.adsabs.harvard.edu/abs/2020arXiv201113456S p

    Song Y., Sohl-Dickstein J., Kingma D. P., Kumar A., Ermon S., Poole B., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2011.13456 , https://ui.adsabs.harvard.edu/abs/2020arXiv201113456S p. arXiv:2011.13456

    Pith/arXiv arXiv doi:10.48550/arxiv.2011.13456 2020

  41. [41]

    arXiv:1503.03757

    Spergel D., et al., 2015, @doi [arXiv e-prints] 10.48550/arXiv.1503.03757 , https://ui.adsabs.harvard.edu/abs/2015arXiv150303757S p. arXiv:1503.03757

    Pith/arXiv arXiv doi:10.48550/arxiv.1503.03757 2015

  42. [42]

    Springel V., White S. D. M., Tormen G., Kauffmann G., 2001, @doi [ ] 10.1046/j.1365-8711.2001.04912.x , https://ui.adsabs.harvard.edu/abs/2001MNRAS.328..726S 328, 726

    arXiv 2001

  43. [43]

    Springel V., et al., 2018, @doi [ ] 10.1093/mnras/stx3304 , https://ui.adsabs.harvard.edu/abs/2018MNRAS.475..676S 475, 676

    doi:10.1093/mnras/stx3304 2018

  44. [44]

    Takada M., et al., 2014, @doi [ ] 10.1093/pasj/pst019 , https://ui.adsabs.harvard.edu/abs/2014PASJ...66R...1T 66, R1

    doi:10.1093/pasj/pst019 2014

  45. [45]

    N., Kaiser L., Polosukhin I., 2017, @doi [arXiv e-prints] 10.48550/arXiv.1706.03762 , https://ui.adsabs.harvard.edu/abs/2017arXiv170603762V p

    Vaswani A., Shazeer N., Parmar N., Uszkoreit J., Jones L., Gomez A. N., Kaiser L., Polosukhin I., 2017, @doi [arXiv e-prints] 10.48550/arXiv.1706.03762 , https://ui.adsabs.harvard.edu/abs/2017arXiv170603762V p. arXiv:1706.03762

    Pith/arXiv arXiv doi:10.48550/arxiv.1706.03762 2017

  46. [46]

    Diffusion Models in 3D Vision: A Survey

    Wang Z., Li D., Wu Y., He T., Bian J., Jiang R., 2024, @doi [arXiv e-prints] 10.48550/arXiv.2410.04738 , https://ui.adsabs.harvard.edu/abs/2024arXiv241004738W p. arXiv:2410.04738

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2410.04738 2024

  47. [47]

    arXiv:2007.14062

    Zaheer M., et al., 2020, @doi [arXiv e-prints] 10.48550/arXiv.2007.14062 , https://ui.adsabs.harvard.edu/abs/2020arXiv200714062Z p. arXiv:2007.14062

    Pith/arXiv arXiv doi:10.48550/arxiv.2007.14062 2020

  48. [48]

    From Dark Matter to Galaxies with Convolutional Networks

    Zhang X., Wang Y., Zhang W., Sun Y., He S., Contardo G., Villaescusa-Navarro F., Ho S., 2019, @doi [arXiv e-prints] 10.48550/arXiv.1902.05965 , https://ui.adsabs.harvard.edu/abs/2019arXiv190205965Z p. arXiv:1902.05965

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1902.05965 2019