Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Contrastively aligning SPHEREx spectra with DESI-LS images substantially improves star–galaxy separation and predicts sub-percent stellar contamination over most of the extragalactic sky.

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 09:20 UTC pith:OYVI27PN

load-bearing objection A solid synthetic proof-of-concept that alignment helps star/galaxy separation for SPHEREx-like data, but the sub-percent contamination forecast is an idealized bound from mocks, not a measured result. the 4 major comments →

arxiv 2607.20797 v1 pith:OYVI27PN submitted 2026-07-22 astro-ph.CO astro-ph.GA

A Multimodal Approach to Star--Galaxy Separation using SPHEREx Spectrophotometry and DESI Legacy Survey Imaging

Kendrick Nguyen , Richard M. Feder , Sean Bruton , Uro\v{s} Seljak This is my paper
classification astro-ph.CO astro-ph.GA
keywords star–galaxy separationcontrastive learningSPHERExDESI Legacy Surveysmultimodal embeddingsstellar contaminationlarge-scale structurephotometric classification
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.

Galaxy clustering analyses aiming at sigma(f_NL) ~ 1 need galaxy samples contaminated by fewer than one star per hundred; stars trace Galactic structure and can fake cosmological power. This paper asks whether forcing SPHEREx near-infrared spectrophotometry and DESI Legacy Survey images into a shared embedding, via contrastive learning, makes stars and galaxies easier to tell apart. It finds yes: aligned image embeddings raise stellar purity from 0.756 to 0.931 and galaxy completeness from 0.835 to 0.961 at a probability threshold of 0.5, and aligned spaces stay nearly as accurate under simple classifiers. Extrapolating to the full SPHEREx footprint with a photo-z precision cut, it forecasts a median stellar contamination of 0.4% when requiring galaxy probability above 0.7. The gain seems to come from alignment reorganizing image embeddings so redshift and near-infrared spectral shape become linearly accessible.

Core claim

The paper establishes that multimodal contrastive alignment reorganizes the embedding space along dimensions better suited to source classification. Using a CLIP-style InfoNCE objective to align a pre-trained image encoder with a spectrum encoder, the authors obtain shared embeddings on which a single XGBoost classifier separates stars from galaxies. On the magnitude-limited COSMOS-like sample, the aligned image embeddings are the strongest representation: stellar purity 0.9308 vs 0.7559 unaligned and galaxy completeness 0.9614 vs 0.8349 at p_gal=0.5, with AUC above 0.992 for every classifier tested. The aligned space is also markedly more linearly separable, with logistic regression within

What carries the argument

A CLIP-style contrastive alignment: an InfoNCE loss with cosine similarity pulls image and spectrum embeddings of the same source together and pushes different sources apart in a shared latent space. Images are encoded by a pre-trained transformer and spectra by a 1D convolutional autoencoder, with multi-head cross-attention projection heads generating 1024-dimensional embeddings; encoders are frozen, only the projection heads train. Downstream, XGBoost classifiers operate on these embeddings, and linear probes measure how accessible redshift and spectral fluxes are. The alignment is what reorganizes the embedding geometry; the paper's key evidence is that this reorganization, not new inform

Load-bearing premise

The load-bearing premise is that the mock SPHEREx spectra and injected stars are as hard to separate as real data; if real spectra, source blending, PSF errors, or noise inhomogeneity make stars and galaxies less separable, the measured gains and the sub-percent contamination forecast will not hold.

What would settle it

Run the same aligned-image classifier on real SPHEREx spectra and DESI-LS cutouts with labels from Gaia and Euclid across 18<z_AB<22.5 and measure stellar purity at p=0.5; if it falls from the predicted ~0.93 toward the unaligned ~0.76, or if the median footprint contamination at p>0.7 exceeds 1%, the paper's forecast is refuted.

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

If this is right

  • A DESI-LS image-only classifier, after alignment, reaches stellar purity 0.931 and galaxy completeness 0.961 on the z<22.5 mock sample, making imaging a much stronger star–galaxy separator than raw spectra alone.
  • Adopting p_gal>0.7 with a sigma_z/(1+z)<0.2 cut yields a predicted median stellar contamination of 0.4% over the SPHEREx extragalactic footprint, with 90% of tiles below 0.9%, meeting sub-percent requirements for sigma(f_NL)~O(1) analyses.
  • Aligned embeddings degrade little when the classifier is simplified: a logistic regression stays within about 0.003 AUC of the full XGBoost model, implying contamination control will be less sensitive to classifier complexity and likely to systematic perturbations.
  • Alignment disproportionately fixes the hardest subpopulations: galaxies at z~0.7–0.9 with 1.4<r-z<2.2, which overlap the stellar locus, are exactly the ones whose image embeddings become linearly separable.
  • Completeness losses from aggressive purity cuts concentrate at low redshift, where SPHEREx samples remain sample-variance limited, so the purity/completeness tradeoff is affordable for clustering analyses.

Where Pith is reading between the lines

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

  • The same contrastive recipe could be applied to other imaging surveys paired with any spectrophotometric catalog; the gain should be largest wherever morphology and broadband colors alone are degenerate, since alignment imports spectral discriminants into the image embedding.
  • Because alignment mostly exposes information already in the imaging, it is a cheaper alternative to adding new bands: no new observations are needed, only a paired spectral catalog at train time. A testable prediction is that gains shrink as survey depth or PSF quality degrades.
  • The mechanism suggests a broader design principle: contrastive alignment can act as a prior that linearizes classification boundaries, so for any high-dimensional astronomical classification task with paired modalities, aligned embeddings should outperform raw embeddings under simple models.
  • A direct test would be to check whether the 1.6-micron H- opacity minimum, fixed in observed wavelength for stars and redshifted for galaxies, is what the aligned image embeddings encode; if so, galaxy sub-samples selected by aligned-image probability should have redshift distributions matching those implied by that feature.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents a CLIP-style contrastive alignment between DESI Legacy Survey images and synthetic SPHEREx spectrophotometry, with the goal of improving star–galaxy separation for the SPHEREx galaxy survey. Using galaxy spectra generated from COSMOS template fits and synthetic stars injected into DESI-LS cutouts, the authors find that aligned embeddings outperform unaligned ones, especially for image-based classification: Table 1 reports galaxy purity 0.9908 and galaxy completeness 0.9614 for aligned image embeddings versus 0.9810 and 0.8349 unaligned. The paper also shows that aligned embeddings are more linearly separable (Section 4.2.3), and it uses photo-z precision cuts to forecast stellar contamination over the full SPHEREx footprint, concluding that sub-percent contamination is achievable with p_gal>0.7 (Section 5.3). The evaluation is entirely synthetic, with caveats acknowledged in Section 5.4.

Significance. If the results transfer to real data, the paper would provide a practical method for improving star–galaxy separation in SPHEREx and Rubin LSST, directly relevant to f_NL science goals. The study is carefully designed: it uses a clean paired image–spectrum construction, compares aligned versus unaligned embeddings under controlled conditions, and includes classifier ablation tests that strengthen the interpretation that alignment restructures the embedding space. The use of publicly available synthetic spectral data (Feder et al. 2023 on Zenodo) and standard open-source tools supports reproducibility. However, the central quantitative claims—the gain in image-based classification and the sub-percent contamination forecast—rest on the realism of the mocks and on the photo-z selection; both need stronger support before the abstract-level claims can be accepted.

major comments (4)
  1. [§5.3, Fig. 12] The sub-percent contamination forecast is derived entirely from synthetic data: galaxies are template SEDs from Feder et al. (2023) and stars are injected at blank-sky positions with a minimum separation of 1.5″ (Section 2.2.3). The caveats in Section 5.4 are qualitative; no quantitative estimate is given for how source blending, spatial noise inhomogeneity, or PSF errors would change η_star. Since the abstract states that contamination 'can be controlled at the sub-percent level across most of the extragalactic sky,' the authors should either validate with real SPHEREx/DESI-LS data (e.g., using Gaia/Euclid cross-checks) or provide a mock-degradation study that quantifies the impact of these effects. Without this, the abstract overstates the strength of the forecast.
  2. [§5.1, Fig. 10] The photo-z precision cut σ_z/(1+z)<0.2 is computed with a template-fitting code whose model grid matches the same template grid used to generate the synthetic galaxy SEDs. This creates a circularity: stars whose spectra resemble template galaxies at moderate redshift may be assigned large photo-z errors and removed preferentially, which is the main lever that suppresses contamination in the forecast. Please test sensitivity to an independent photo-z method (e.g., a different template set, a neural photo-z estimator, or real stellar SEDs) and report how the contamination fractions in §5.3 change. This is a load-bearing point because a factor-of-two change in the high-redshift stellar tail would materially alter the conclusion.
  3. [Table 1, §4.2.2] The classification metrics are reported without error bars, even though the 5-fold cross-validation procedure yields a distribution. The image-based improvement claims rely on differences of ~17 percentage points in stellar purity and ~13 percentage points in galaxy completeness; reporting fold-to-fold scatter or bootstrap uncertainties is necessary to confirm that the improvement is not driven by a particular split. The same applies to the AUC values in Figure 8 and the completeness/purity curves in Figures 6 and 11.
  4. [§4.2.1 and §5.3] The stellar-density correction is hand-calibrated to a ~50% overprediction near COSMOS and then extrapolated to the full sky using Gaia star counts. This correction is applied as a fixed factor, and its uncertainty is not propagated into the contamination forecast. Moreover, as the authors note in Section 5.4, Gaia may not trace the same stellar populations that leak into galaxy samples. Please provide a sensitivity analysis—e.g., varying the correction factor by ±50%—and show how the median and the 90/99th-percentile contamination fractions in Figure 12 change.
minor comments (5)
  1. [§4.2.4, Table 2] The subsample R² values are computed on the set of sources that are misclassified before alignment and correctly classified after. The sample size and the restricted range of this subset should be stated; R² on small, selected subsets can be noisy and should be interpreted with caution.
  2. [§5.3] The sentence 'η_star has a median of 1.3%, with 90% of the footprint having η_star<2.2% for 90% of HEALPix tiles and <3.6% for 99% of tiles' is confusing. Please rephrase, e.g., 'the median contamination fraction is 1.3%; 90% of the footprint has η_star<2.2% and 99% has η_star<3.6%.'
  3. [§3.2] The text describes the 102-channel SPHEREx data as 'photometry' in the sentence 'mapping the 102-channel photometry to a 6-dimensional latent space.' These are low-resolution spectral channels, not broadband photometry; please adjust the terminology for clarity.
  4. [Figure 3] The caption contains a typo: 'In the top bottom' should read 'In the bottom row.'
  5. [§2.2.3] The definition of the blank-sky injection separation θ_min=1.5″ is clear, but the paper does not state how many injected stars were rejected for lack of valid blank-sky positions. A sentence on the success rate would help assess potential selection biases.

Circularity Check

0 steps flagged

No significant circularity: the central classification and contamination results are measured on held-out synthetic data and do not reduce to fitted inputs or same-author self-citations.

full rationale

The derivation chain is self-contained and its main quantitative claims are not equivalent to the inputs by construction. The contrastive alignment (Eq. 1) is trained on image–spectrum pairs without using star/galaxy labels; labels enter only at the XGBoost stage, and Table 1 metrics are computed on held-out validation folds (5-fold CV, §4.2.2). The image-based classification gains therefore measure genuine generalization within the mock distribution, not the training loss itself. The contamination forecast (§5.3) rescales measured per-object false-positive rates using Gaia stellar-density maps; the only calibrated parameter is the ~50% stellar-density correction (§4.2.1), which the paper explicitly labels approximate ('We caution that this correction is approximate and defer a full comparison of simulations and data to future work') and which changes normalization, not the mechanism. The one same-author citation, Feder et al. (2023), supplies the synthetic galaxy SEDs; it is public on Zenodo, its template and emission-line assumptions are stated, and the stellar SEDs come from an independent BaSeL/pystellib model, so the classification is not forced by that citation. The photo-z step (§5.1) uses the Stickley et al. code with a Feder et al. grid, but this is a standard mock-reality test: the stellar spectra are not drawn from that galaxy grid, and the paper itself calls the high-redshift stellar solutions 'fictitious,' showing they are an output, not an input. The paper's own §5.4 and Conclusion explicitly identify the real-data validation gap ('Validating and extending these methods will require testing on real SPHEREx spectra and imaging'), which is an external-validity limitation, not circularity. No fitted parameter is renamed as a prediction, and no 'uniqueness' theorem is imported from the authors' prior work.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

No new physical entities are introduced. The paper's central claim rests instead on simulation fidelity: synthetic SPHEREx spectra, synthetic stars, and injected images must reproduce the real discriminative structure of the data. The free parameters are mostly experimental design choices, with the stellar-density correction being the one that materially moves the survey forecast.

free parameters (6)
  • mock stellar density correction factor = ~1.5 (mocks overpredict Gaia star counts by ~50% near COSMOS)
    Applied in Section 4.2.1 to reweight the star sample; directly sets the normalization of all purity, FPR, and contamination numbers, including the sub-percent forecast.
  • uniform E(B-V)=0.02 reddening for synthetic stars = 0.02
    Chosen to match COSMOS extinction (Section 2.2.3); affects stellar colors and therefore classification difficulty, though it is not fitted in this paper.
  • blank-sky injection minimum separation theta_min=1.5 arcsec = 1.5 arcsec
    Synthetic stars are placed at least 1.5 arcsec from catalogued sources and other injected stars (Section 2.2.3); this avoids catastrophic blends and makes the image classification task easier than real survey data.
  • CLIP temperature tau = 15.0
    Fixed from AstroCLIP (Section 3.3); a hyperparameter that controls the sharpness of the contrastive loss and hence the geometry of the aligned space.
  • autoencoder bottleneck dimension = 6 (12 tested)
    The spectrum encoder uses a 6-dimensional latent bottleneck (Section 3.2); the authors show 6 versus 12 gives nearly identical downstream classification, so this is not load-bearing.
  • XGBoost default hyperparameters = 200 trees, max depth 6
    Chosen by hand (Section 3.4); the ablation with reduced trees/depth shows the aligned embeddings are insensitive to this choice, so it is not load-bearing.
axioms (6)
  • domain assumption SPHEREx synthetic spectra from Feder et al. (2023) realistically emulate real SPHEREx spectrophotometry, including photometric noise.
    All spectrum-based classification, photo-z estimates, and contamination forecasts use these mocks; if the simulated noise or SEDs are too optimistic, the headline contamination numbers will not transfer. The paper does not validate against real SPHEREx data.
  • domain assumption pystellib/BaSeL synthetic stellar SEDs, Galaxia spatial distribution, and PSF injection accurately represent the real stellar population seen by DESI-LS and SPHEREx.
    The star sample, including its density and colors, drives stellar contamination estimates; the ~50% density correction and uniform extinction are patches for known mock-data mismatches acknowledged in Section 2.2.3.
  • domain assumption The AstroDINO image encoder, pretrained on DESI-LS galaxies, retains enough information for star-galaxy separation after alignment.
    The image branch is frozen during alignment, so all image results inherit the quality and biases of the AstroDINO representation; the paper does not retrain or validate this encoder on stars.
  • domain assumption The template-fitting photo-z code of Stickley et al. (2016) produces usable p(z) distributions for both stars and galaxies when applied to synthetic SPHEREx spectra.
    The redshift-error selections in Section 5, which remove most stellar contamination, depend entirely on these photo-z estimates for the mocks.
  • standard math The InfoNCE/CLIP contrastive objective is a valid proxy for maximizing mutual information between paired image and spectrum embeddings.
    Standard result from van den Oord et al. (2018) and Radford et al. (2021); the paper uses it as the alignment engine without re-deriving it.
  • domain assumption COSMOS 2020 templates plus the semi-empirical emission-line model of Feder et al. (2023) cover the true diversity of galaxies at z<2 in the SPHEREx sample.
    The galaxy side of the training set is generated from these templates; missing galaxy types or emission-line behaviors would change the separability measured here.

pith-pipeline@v1.3.0-alltime-deepseek · 21868 in / 11190 out tokens · 113346 ms · 2026-08-01T09:20:19.167741+00:00 · methodology

0 comments
read the original abstract

Stellar contamination is a critical systematic for increasingly precise large-scale structure analyses from ongoing and next-generation surveys. Experiments targeting constraints on local primordial non-Gaussianity with $\sigma(f_{\rm NL}^{\rm loc}) \sim \mathcal{O}(1)$ demand sub-percent stellar contamination rates to avoid misidentifying spurious large-scale power induced by Galactic structure as true cosmological signal. In this work, we explore the use of multimodal models for star--galaxy separation, harnessing the information from both optical broad-band imaging data and SPHEREx near-infrared low-resolution spectrophotometry. The two modalities are integrated using contrastive learning, which projects image- and spectrum-based embeddings into a shared latent space. We find that classifiers trained on these transformed representations outperform those trained on the original embeddings and show less performance degradation when simpler classifiers are used. These results suggest that multimodal alignment organizes the embedding space along dimensions that are better suited to source classification. The improvement is particularly strong for image-based classification, which we connect to increased predictability of highly-discriminative infrared spectral features from the transformed image embeddings. Applying redshift error-based selections and extrapolating to the full SPHEREx footprint, we demonstrate that stellar contamination can be controlled at the sub-percent level across most of the extragalactic sky, with completeness tradeoffs largely confined to low redshift. Our work highlights the utility of multimodal methods for modern galaxy surveys such as SPHEREx and $\textit{Rubin}$ LSST.

Figures

Figures reproduced from arXiv: 2607.20797 by Kendrick Nguyen, Richard M. Feder, Sean Bruton, Uro\v{s} Seljak.

Figure 1
Figure 1. Figure 1: Left: Distribution of object counts for stars and galaxies as a function of z-band magnitude. Right: Color-color diagram of galaxies and stars. Contours are spaced at logarithmic intervals in source density. Stars and galaxies exhibit substantial overlap in optical broadband color-color space, with approximately 65% of galaxies and 69% of stars falling within the other’s 95% density region. The alignment o… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of alignment framework. The encoders output representations of their respective data, which are then projected into a shared embedding space. A contrastive loss encourages the projection heads to push image-spectrum pairs of the same object together and mismatched pairs apart. trum is represented as a sequence of localized vectors rather than a single global vector. This representation is analogo… view at source ↗
Figure 3
Figure 3. Figure 3: UMAP visualization of encoder embedding spaces. Left: unaligned spectrum encoder. Middle: unaligned image encoder. Right: spectrum encoder after contrastive alignment with the image encoder. In the top row, point are colored by their ground-truth labels, with galaxies shown in yellow and stars in purple. In the top bottom, they are colored by the galaxy probability assigned by the XGBoost classifier. The a… view at source ↗
Figure 4
Figure 4. Figure 4: Examples of spectra with high, medium, and low class probabilities assigned by the XGBoost classifier trained on embeddings of the aligned spectrum encoder. The central 40 × 40 pixels of the corresponding z-band images are also shown, together with the class probabilities assigned by an XGBoost classifier trained on embeddings of the aligned image encoder. These examples illustrate how sources with uncerta… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of galaxies for which the spectrum￾based classifier assigns a high galaxy probability while the image-based classifier assigns a low galaxy probability. Im￾ages show the central 40 × 40 pixels of the full cutout for easier visualization. Probabilities are produced by XGBoost classifiers trained on unaligned spectrum and image embed￾dings, respectively. with the classifier trained on aligned image … view at source ↗
Figure 6
Figure 6. Figure 6: Galaxy purity and completeness of the XGBoost classifier using embeddings from the unaligned spectrum encoder (left), classification performed directly on mock SPHEREx spectra (middle), embeddings from the aligned spectrum encoder (right). We show results for classifiers evaluated on both the fiducial noise level test data (solid curves) and spectra with √ 2× higher flux uncertainties (dashed). 1 2 3 4 5 W… view at source ↗
Figure 7
Figure 7. Figure 7: Feature importance as a function of observed wavelength, for a XGBoost classifier trained directly on spec￾tra (§4). The feature importance peaks at channels near 1.1, 1.6, and 2.9 µm, suggesting that the discriminative informa￾tion is concentrated near a small number of spectral regions. splits using that feature, computed over all nodes and trees in the ensemble. As such, features with high im￾portance a… view at source ↗
Figure 8
Figure 8. Figure 8: ROC curves for the XGBoost classifier. Left: embeddings from the unaligned spectrum encoder. Middle: classifi￾cation performed directly on spectra. Right: embeddings from the aligned spectrum encoder. Curves are shown for our default XGBoost hyperparameters (max depth=6, n estimators=200), a reduced XGBoost model (max depth=3, n estimators=50), a logistic regression classifier, and a K-nearest neighbors cl… view at source ↗
Figure 9
Figure 9. Figure 9: Properties of sources whose image-based classifications improve after contrastive alignment. Left: Optical color￾color distribution of the full validation set, with the 70 sources showing the largest increase in classification probability for their true class after alignment highlighted. Right: Redshift distribution of all validation-set galaxies compared to galaxies that are misclassified before alignment… view at source ↗
Figure 10
Figure 10. Figure 10: Impact of classification- and photo-z-based selections on simulated galaxies (top) and stars (bottom). In each panel we include the distribution of redshifts estimated using the template fitting code. Note that stars are plotted on a logarithmic y-axis scale. based classifier selections applied (with p thresh gal = 0.5). Across the three bins, the two modalities are comple￾mentary for galaxy recovery. In … view at source ↗
Figure 11
Figure 11. Figure 11: Impact of different classification probability thresholds for galaxy completeness (blue) and false positive rate (red) as a function of redshift, where here we define metrics for a sample with zAB < 22.5 and σz/(1+z) < 0.2 cuts applied. These false positive rate estimates assume stellar density of the COSMOS field. In [PITH_FULL_IMAGE:figures/full_fig_p016_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Forecasted stellar contamination fraction as a function of sky position, for different classification probability thresholds. In each case we include a notional mask that excludes regions in the Galactic plane, as well as additional masks for the LMC, SMC, etc. The black cross indicates the position of the COSMOS field. These results extrapolate classification performance from the aligned, spectrum-based … view at source ↗
Figure 13
Figure 13. Figure 13: UMAP visualization of encoder embedding spaces. Top: aligned image embeddings of stars and galaxies. Bottom: aligned spectrum embeddings of stars and galaxies. While the contrastive objective encourages spectra and imaging of the same sources to reside at the same position in the shared embedding space, some discrepancies remain (see text). In [PITH_FULL_IMAGE:figures/full_fig_p019_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: UMAP visualization of galaxy embeddings from the aligned spectrum encoder. Left: points colored by redshift. Middle: points colored by stellar mass. Right: points colored by Hα equivalent width. The embedding exhibits clear organization with redshift and stellar mass, while the relationship with Hα equivalent width appears noisier, though some structure remains visible. While alignment may improve the org… 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

37 extracted references · 4 canonical work pages · 1 internal anchor

  1. [1]

    P., Crill, B

    Akeson, R., Dubois-Felsmann, G. P., Crill, B. P., et al. 2025, arXiv e-prints, arXiv:2511.15823, doi: 10.48550/arXiv.2511.15823 Astropy Collaboration, Robitaille, T. P., Tollerud, E. J., et al. 2013, A&A, 558, A33, doi: 10.1051/0004-6361/201322068

    Pith/arXiv arXiv doi:10.48550/arxiv.2511.15823 2025

  2. [2]

    2014, doi: 10.48550/arXiv.1403.5237

    Benitez, N., Dupke, R., Moles, M., et al. 2014, doi: 10.48550/arXiv.1403.5237

    Pith/arXiv arXiv doi:10.48550/arxiv.1403.5237 2014

  3. [3]

    2011, in Astronomical Society of the Pacific Conference Series, Vol

    Bertin, E. 2011, in Astronomical Society of the Pacific Conference Series, Vol. 442, Astronomical Data Analysis Software and Systems XX, ed. I. N. Evans, A. Accomazzi, D. J. Mink, & A. H. Rots, 435

    2011

  4. [4]

    J., Aboobaker, A

    Bock, J. J., Aboobaker, A. M., Adamo, J., et al. 2025, arXiv e-prints, arXiv:2511.02985, doi: 10.48550/arXiv.2511.02985 21

    doi:10.48550/arxiv.2511.02985 2025

  5. [5]

    2016, arXiv e-prints, arXiv:1603.02754, doi: 10.48550/arXiv.1603.02754

    Chen, T., & Guestrin, C. 2016, arXiv e-prints, arXiv:1603.02754, doi: 10.48550/arXiv.1603.02754

    Pith/arXiv arXiv doi:10.48550/arxiv.1603.02754 2016

  6. [6]

    O., Scaife, A

    Clarke, A. O., Scaife, A. M. M., Greenhalgh, R., & Griguta, V. 2020, A&A, 639, A84, doi: 10.1051/0004-6361/201936770 de Jong, J. T. A., Verdoes Kleijn, G. A., Kuijken, K. H., &

    work page doi:10.1051/0004-6361/201936770 2020

  7. [7]

    Valentijn, E. A. 2013, Experimental Astronomy, 35, 25, doi: 10.1007/s10686-012-9306-1 DESI Collaboration, Aghamousa, A., Aguilar, J., et al. 2016, arXiv e-prints, arXiv:1611.00036, doi: 10.48550/arXiv.1611.00036 DESI Legacy Surveys. 2023, Galactic Extinction Coefficients, https://www.legacysurvey.org/dr10/ catalogs/#galactic-extinction-coefficients

    Pith/arXiv arXiv doi:10.48550/arxiv.1611.00036 2013

  8. [8]

    2019, The Astronomical Journal, 157, 168, doi: 10.3847/1538-3881/ab089d

    Dey, A., et al. 2019, The Astronomical Journal, 157, 168, doi: 10.3847/1538-3881/ab089d

    doi:10.3847/1538-3881/ab089d 2019

  9. [9]

    M., Masters, D

    Feder, R. M., Masters, D. C., Lee, B., et al. 2023, arXiv e-prints, arXiv:2312.04636, doi: 10.48550/arXiv.2312.04636

    Pith/arXiv arXiv doi:10.48550/arxiv.2312.04636 2023

  10. [10]

    Fitzpatrick, E. L. 1999, PASP, 111, 63 Gagn´ e, J., Faherty, J. K., Ruiz Diaz, A., et al. 2026, arXiv e-prints, arXiv:2604.22012, doi: 10.48550/arXiv.2604.22012 Gaia Collaboration, Vallenari, A., Brown, A. G. A., et al. 2023, A&A, 674, A1, doi: 10.1051/0004-6361/202243940

    Pith/arXiv arXiv doi:10.48550/arxiv.2604.22012 1999

  11. [11]

    Girardi, L., Groenewegen, M. A. T., Hatziminaoglou, E., & da Costa, L. 2005, A&A, 436, 895, doi: 10.1051/0004-6361:20042352 Gopalakrishnan Nair, N., Gedara Chaminda Bandara, W., & Patel, V. M. 2022, arXiv e-prints, arXiv:2206.05039, doi: 10.48550/arXiv.2206.05039

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2206.05039 2005

  12. [12]

    R., Millman, K

    Harris, C. R., Millman, K. J., van der Walt, S. J., et al. 2020, Nature, 585, 357, doi: 10.1038/s41586-020-2649-2

    doi:10.1038/s41586-020-2649-2 2020

  13. [13]

    J., Cheng, Y.-T., et al

    Huai, Z., Bock, J. J., Cheng, Y.-T., et al. 2025, arXiv e-prints, arXiv:2510.01410, doi: 10.48550/arXiv.2510.01410

    doi:10.48550/arxiv.2510.01410 2025

  14. [14]

    Hunter, J. D. 2007, Computing in Science & Engineering, 9, 90, doi: 10.1109/MCSE.2007.55 Ivezi´ c,ˇZ., Kahn, S. M., Tyson, J. A., et al. 2019, ApJ, 873, 111, doi: 10.3847/1538-4357/ab042c

    doi:10.1109/mcse.2007.55 2007

  15. [15]

    P., Vieira dos Santos, G., Marra, V., et al

    Jeakel, A. P., Vieira dos Santos, G., Marra, V., et al. 2025, arXiv e-prints, arXiv:2511.20524. https://arxiv.org/abs/2511.20524

    arXiv 2025

  16. [16]

    W., & Mykytyn, D

    Lang, D., Hogg, D. W., & Mykytyn, D. 2016, The Tractor: Probabilistic astronomical source detection and measurement, Astrophysics Source Code Library, record ascl:1604.008. http://ascl.net/1604.008

    2016

  17. [17]

    1997, A&AS, 125, 229, doi: 10.1051/aas:1997373

    Lejeune, T., Cuisinier, F., & Buser, R. 1997, A&AS, 125, 229, doi: 10.1051/aas:1997373

    doi:10.1051/aas:1997373 1997

  18. [18]

    2019, arXiv preprint arXiv:1711.05101

    Loshchilov, I., & Hutter, F. 2019, arXiv preprint arXiv:1711.05101

    Pith/arXiv arXiv 2019

  19. [19]

    2020, in Proceedings of Machine Learning Research, Vol

    McAllester, D., & Stratos, K. 2020, in Proceedings of Machine Learning Research, Vol. 108, Proceedings of the Twenty Third International Conference on Artificial Intelligence and Statistics, ed. S. Chiappa & R. Calandra (PMLR), 875–884. https://proceedings.mlr.press/v108/mcallester20a.html

    2020

  20. [20]

    2018, arXiv preprint arXiv:1802.03426

    McInnes, L., Healy, J., & Melville, J. 2018, arXiv preprint arXiv:1802.03426

    Pith/arXiv arXiv 2018

  21. [21]

    2024, MNRAS, 531, 4990, doi: 10.1093/mnras/stae1450

    Parker, L., Lanusse, F., Golkar, S., et al. 2024, MNRAS, 531, 4990, doi: 10.1093/mnras/stae1450

    doi:10.1093/mnras/stae1450 2024

  22. [22]

    2025, arXiv e-prints, doi: 10.48550/arXiv.2510.17960

    Parker, L., Lanusse, F., Shen, J., et al. 2025, arXiv e-prints, doi: 10.48550/arXiv.2510.17960

    doi:10.48550/arxiv.2510.17960 2025

  23. [23]

    2019, in Advances in Neural Information Processing Systems 32 (Curran Associates, Inc.), 8024–8035

    Paszke, A., Gross, S., et al. 2019, in Advances in Neural Information Processing Systems 32 (Curran Associates, Inc.), 8024–8035. http://nips.cc

    2019

  24. [24]

    W., Hallacy, C., et al

    Radford, A., Kim, J. W., Hallacy, C., et al. 2021, arXiv e-prints, arXiv:2103.00020, doi: 10.48550/arXiv.2103.00020

    Pith/arXiv arXiv doi:10.48550/arxiv.2103.00020 2021

  25. [25]

    2026, arXiv e-prints, arXiv:2607.00543

    Rustamkulov, Z., Kirkpatrick, J., Akeson, R., et al. 2026, arXiv e-prints, arXiv:2607.00543. https://arxiv.org/abs/2607.00543

    Pith/arXiv arXiv 2026

  26. [26]

    2002, The Astronomical Journal, 124, 3050, doi: 10.1086/344682

    Sawicki, M. 2002, The Astronomical Journal, 124, 3050, doi: 10.1086/344682

    doi:10.1086/344682 2002

  27. [27]

    2007, The Astrophysical Journal Supplement Series, 172, 1, doi: 10.1086/516585

    Scoville, N., Aussel, H., Brusa, M., et al. 2007, The Astrophysical Journal Supplement Series, 172, 1, doi: 10.1086/516585

    doi:10.1086/516585 2007

  28. [28]

    2011, ApJ, 730, 3, doi: 10.1088/0004-637X/730/1/3

    Binney, J. 2011, ApJ, 730, 3, doi: 10.1088/0004-637X/730/1/3

    doi:10.1088/0004-637x/730/1/3 2011

  29. [29]

    R., Capak, P., Masters, D., et al

    Stickley, N. R., Capak, P., Masters, D., et al. 2016, arXiv e-prints, arXiv:1606.06374, doi: 10.48550/arXiv.1606.06374 The Dark Energy Survey Collaboration. 2005, arXiv e-prints, astro, doi: 10.48550/arXiv.astro-ph/0510346 The Multimodal Universe Collaboration, Audenaert, J.,

    Pith/arXiv arXiv doi:10.48550/arxiv.1606.06374 2016

  30. [30]

    2024, arXiv e-prints, arXiv:2412.02527, doi: 10.48550/arXiv.2412.02527 van den Oord, A., Li, Y., & Vinyals, O

    Bowles, M., et al. 2024, arXiv e-prints, arXiv:2412.02527, doi: 10.48550/arXiv.2412.02527 van den Oord, A., Li, Y., & Vinyals, O. 2018, CoRR, abs/1807.03748. https://arxiv.org/abs/1807.03748

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.02527 2024

  31. [31]

    E., et al

    Virtanen, P., Gommers, R., Oliphant, T. E., et al. 2020, Nature Methods, 17, 261, doi: 10.1038/s41592-019-0686-2

    doi:10.1038/s41592-019-0686-2 2020

  32. [32]

    R., Kauffmann, O

    Weaver, J. R., Kauffmann, O. B., Ilbert, O., et al. 2022, ApJS, 258, 11, doi: 10.3847/1538-4365/ac3078

    doi:10.3847/1538-4365/ac3078 2022

  33. [33]

    2021, MNRAS, 503, 5061, doi: 10.1093/mnras/stab709

    Weaverdyck, N., & Huterer, D. 2021, MNRAS, 503, 5061, doi: 10.1093/mnras/stab709

    work page doi:10.1093/mnras/stab709 2021

  34. [34]

    Y., Gebhardt, H

    Wen, R. Y., Gebhardt, H. S. G., Heinrich, C., & Dor´ e, O. 2025, PhRvD, 112, 063518, doi: 10.1103/hqc3-tbxc

    work page doi:10.1103/hqc3-tbxc 2025

  35. [35]

    Yang, Y. et al. 2025, in prep. 22

    2025

  36. [36]

    2026, ApJ, 998, 189, doi: 10.3847/1538-4357/ae2c7e

    Zhao, X., Huang, Y., Xue, G., et al. 2026, ApJ, 998, 189, doi: 10.3847/1538-4357/ae2c7e

    doi:10.3847/1538-4357/ae2c7e 2026

  37. [37]

    A., et al

    Zhou, R., Dey, B., Newman, J. A., et al. 2023, AJ, 165, 58, doi: 10.3847/1538-3881/aca5fb

    doi:10.3847/1538-3881/aca5fb 2023