Pith. sign in

REVIEW 1 major objections 7 minor

From DESI to Euclid: A Generative Bridge to Improve Measurements of Galaxy Structure

T0 review · 1 major / 7 minor · reviewed 2026-07-09 · glm-5.2

Pith's one-line read Generative model sharpens ground-based galaxy images, removing size bias

desk verdict Generative DESI-to-Euclid image translation works well on the test set; the released catalog's out-of-distribution generalization is untested but honestly framed. read the letter →

arxiv 2607.06891 v2 pith:V5HQLKBS submitted 2026-07-08 astro-ph.GA astro-ph.IM

classification astro-ph.GAastro-ph.IM
keywords predictionsdesieuclidgalaxymeasurementsbiasbandbiases
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Ground-based telescope images of galaxies are blurred by Earth's atmosphere, which systematically distorts measurements of galaxy size and structure. This distortion is size-dependent: smaller galaxies are affected more, corrupting the scaling relations astronomers use to trace how galaxies grow. Space telescopes solve the problem but cover only a sliver of the sky. This paper trains a generative diffusion model called an Image-to-Image Schrödinger Bridge on the small overlap between ground-based DESI imaging and space-based Euclid imaging, learning to translate blurry DESI images into sharp Euclid-quality images. The key mechanism is a stochastic bridge: the model learns to reverse a gradual blurring process that connects a sharp Euclid image to its blurry DESI counterpart, recovering structure that the atmosphere destroyed. The authors show, using Fourier-domain analysis, that the recovered structure is genuinely constrained by the input data down to 0.37 arcseconds—a 3.8-fold improvement over the 1.41-arcsecond DESI baseline—rather than being fabricated by the model's learned prior. At this recovery level, the systematic biases in three standard structural parameters (Petrosian radius, Sérsic radius, and Sérsic index) are essentially eliminated, removing the size-dependent distortion that plagues ground-based measurements. The authors release their translations across the full Euclid DR1 footprint as a dataset called E-BGS, which can be validated once Euclid's real images become public.

What carries the argument

The Image-to-Image Schrödinger Bridge (I2SB): a diffusion model that defines a stochastic path between two fixed image endpoints—a sharp Euclid image and its blurry DESI counterpart. The forward process gradually blurs the Euclid image into the DESI one; a neural network learns to reverse this path, recovering sharp structure from the blurry input. The bridge is stochastic (not a fixed interpolation), which lets it express the one-to-many nature of recovering fine structure from degraded data. A Fourier Ring Correlation test checks whether recovered structure matches the ground truth in phase, not just in power, distinguishing genuine data-constrained recovery from prior-driven invention.

What would settle it

If the released E-BGS predictions over the Euclid DR1 footprint, once compared against real Euclid DR1 images, show systematic biases in structural parameters that depend on galaxy size, type, or local noise conditions—or if the Fourier Ring Correlation with real Euclid data falls below the 0.5 threshold at scales coarser than 0.37 arcseconds—the claim of unbiased, data-constrained recovery would fail.

Watch

Extended reading notes

Core claim

A bridge diffusion model trained on 63 square degrees of real DESI–Euclid image pairs can translate blurry ground-based galaxy images into near-space-based resolution well enough to remove the systematic, size-dependent biases in structural parameters. The recovery is data-constrained—confirmed by phase coherence in the Fourier domain down to 0.37 arcseconds—rather than prior-driven fabrication, and the de-biasing holds across the Petrosian radius, Sérsic radius, and Sérsic index simultaneously.

Load-bearing premise

The model is trained on 63 square degrees of Euclid Deep Fields and applied to the full Euclid DR1 footprint, but the deep fields may have different noise properties, depth, or source distributions than the wide survey. The paper does not test whether the model generalizes across genuinely distinct sky regions beyond the spatially disjoint test set carved from the same deep-field data.

Editorial extensions

If this is right

  • Population-level studies of galaxy structure—such as the mass–size relation—can proceed at near-Euclid resolution across the full DESI BGS footprint before Euclid imaging is available, using the released E-BGS dataset.
  • The phase-coherence validation method (Fourier Ring Correlation) provides a general framework for testing whether any generative image-translation model in astronomy recovers real structure versus fabricating plausible-looking but unconstrained output.
  • If the E-BGS predictions are confirmed against real Euclid DR1 data, the same bridge-diffusion approach could be applied to other ground-to-space translation problems in astronomy, such as HSC-to-Hubble or LSST-to-Roman.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The claim that bias removal matters more than scatter for population studies is well-taken, but it implicitly assumes that the residual scatter is uncorrelated with galaxy properties. If the model's plausible-but-wrong predictions for unresolved central profiles cluster systematically at certain masses or redshifts, the scatter could still bias scaling-relation slopes even if the mean bias is zero
  • The model's trusted scale of 0.37 arcseconds is set by where the Fourier Ring Correlation drops to 0.5, but this is a population-level statistic. Individual galaxies—especially the smallest or most compact ones—may have substantially worse recovery, meaning the de-biasing may not be uniform across the full BGS sample.
  • The approach could be extended to recover structural parameters that are currently impossible to measure from the ground at all, such as bulge-to-disk decomposition or central velocity dispersion proxies, if the model can be shown to preserve the relevant small-scale information faithfully.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 7 minor

Summary. This paper presents an Image-to-Image Schrödinger Bridge (I2SB) model that translates DESI Legacy Imaging Surveys r- and z-band images of Bright Galaxy Survey (BGS) targets into Euclid VIS-resolution images. The model is trained on 63.1 deg² of observed DESI–Euclid Q1 overlap and validated on a spatially disjoint test region. The authors demonstrate, via Fourier Ring Correlation (FRC), that recovered structure is phase-coherent with Euclid ground truth down to 0.37″ (from 1.41″ in DESI r), and that structural parameters (Petrosian radius, Sérsic radius, Sérsic index) measured from the predictions are substantially de-biased relative to DESI measurements. The authors release predictions over the full Euclid DR1 footprint as E-BGS, framing the release as a blind prediction to be validated once DR1 is public. The methodology is sound, the validation framework is well-designed, and the FRC-based test for data-constrained versus prior-invented structure is a genuine methodological strength.

Significance. The paper addresses a real and timely problem: the majority of BGS galaxies lack space-based imaging, and seeing-induced biases in structural parameters cannot be fully removed by post-hoc correction once galaxies are under-resolved. The generative approach, validated against two well-motivated criteria (data-constrained structure and unbiased parameters), is a credible contribution. Particular strengths include: (1) the FRC phase-coherence test, which directly addresses whether recovered structure is genuinely data-constrained rather than prior-fabricated; (2) the transparent framing of the DR1 release as a falsifiable prediction; (3) the release of both code and predictions, enabling community validation; and (4) the honest acknowledgment that scatter is not reduced, with a clear argument that bias removal is what matters for population-level scaling relations. The work is a meaningful step toward enabling unbiased structural measurements ahead of full Euclid coverage.

major comments (1)
  1. §2.3, §4: The most load-bearing concern is the gap between the validated domain and the released product. The model is trained and tested on data drawn entirely from the three Euclid Deep Fields (Q1), which, while at nominal Wide Survey depth, may differ from the DR1 wide-survey footprint in stellar density, galactic extinction, background characteristics, or source distributions. The released E-BGS catalog covers the full DR1 footprint, but no test of cross-field generalization is presented. The authors are transparent that this is a prediction to be validated later, which is appropriate. However, a leave-one-field-out cross-validation—training on two of the three deep fields and testing on the third—would directly assess whether field-specific features have been learned and would substantially strengthen the claim that the released predictions are expected to be unbiased. If this is in
minor comments (7)
  1. §4.3: The fraction of the full BGS sample retained after the Sérsic-fit quality cut (χ²_ν < 3 on the GT) is not reported. Since the de-biasing results for Re and n are presented only on this clean subset, the reader needs to know what fraction of the population it represents to assess generalizability of the structural-parameter claims to the full BGS.
  2. Figure 4b: The three FWHM-binned FRC curves are described in the caption but are difficult to distinguish in the figure. Consider using more distinct line styles or a separate panel to make the ~0.04″ variation across PSF bins clearly visible.
  3. §3.2: The statement that the per-band stretch ensures 'each step only adds structure toward the Euclid endpoint and never carries the DESI structure across' is a strong claim about the mechanism. A brief clarification of how the stretch parameters achieve this (beyond the Appendix A reference) would help the reader evaluate this design choice.
  4. Abstract: 'approximately 3.8-fold improvement' could be stated more precisely as '~3.8×' for consistency with the body text, where the symbol is used.
  5. §2.1: The statement that the g-band 'lies almost entirely below the Euclid VIS bandpass' could benefit from a quantitative overlap fraction for precision.
  6. Table 1 caption: 'All values in native flux units' should specify the unit system (e.g., nanomaggies) for reproducibility.
  7. Appendix A.5: The Optuna objective function is described as minimizing 'the bias and scatter of its recovered Sérsic and photometric parameters,' but the exact functional form of this score is not given. A brief specification would aid reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: validation against independent Euclid ground truth on a spatially disjoint test set, with standard diagnostics

full rationale

The paper's derivation chain is self-contained and non-circular. The I2SB model is trained on observed DESI–Euclid image pairs over the Q1 overlap (63.1 deg²), then validated on a spatially disjoint test set (Section 2.3: 'The test set is the region where Q1 and DESI DR1 overlap; the remaining, spatially disjoint pairs form the training and validation sets'). The two validation criteria are each measured against an independent ground truth: (1) the FRC trusted-scale analysis (Section 4.2) uses the standard FRC=0.5 threshold from van Heel & Schatz 2005, comparing E-BGS predictions to Euclid GT images—no fitted parameter is renamed as a prediction; (2) the structural-parameter recovery (Section 4.3) measures Petrosian radius, Sérsic radius, and Sérsic index independently on GT, E-BGS, and DESI images using statmorph, with residuals defined as GT minus method. The per-band stretch parameters (Table 1) are dataset-level preprocessing fixed before training, not fitted to the target result. Model selection (Appendix A.5) tunes hyperparameters on a held-out validation subset, then evaluates on the separate test set—standard train/validate/test methodology. The released E-BGS catalog over the DR1 footprint is explicitly framed as a prediction 'to be blindly validated once DR1 is public,' not as a validated result. The I2SB framework is cited from external authors (Liu et al. 2023), not self-cited. No step in the chain reduces to its own inputs by construction.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or postulated objects. All free parameters are either I2SB framework defaults, Optuna-tuned hyperparameters, or dataset-level preprocessing constants. The axioms are standard mathematical results or domain-specific assumptions about survey representativeness and fitting validity.

free parameters (8)
  • beta_min = 1e-10
    Noise schedule lower bound, set by the I2SB framework defaults.
  • beta_max = 0.15
    Noise schedule upper bound, tuned via Optuna over 25 trials.
  • T = 500
    Number of schedule steps, tuned via Optuna.
  • K (reverse steps) = 10
    Number of inference sub-samples, tuned via Optuna.
  • UNet base channels = 64
    Network width, tuned via Optuna.
  • Channel multipliers = (1,2,4,4)
    Network depth scaling, tuned via Optuna.
  • Per-band stretch parameters (L_b, U_b, m_b, sigma_b) = See Table 1
    Dataset-level preprocessing parameters fixed before training to map native flux to [-1,1].
  • Sersic fit bounds = 0.01<=Re<=100, 0.25<=n<=8
    Fitting constraints for statmorph, chosen by the authors.
assumptions (4)
  • standard math The I2SB Schrodinger Bridge framework (Liu et al. 2023) provides a valid stochastic interpolation between two image distributions.
    Invoked in Section 3.1 as the mathematical foundation for the forward/reverse process.
  • domain assumption The FRC=0.5 threshold is a valid criterion for determining the trusted spatial scale of recovered structure.
    Used in Section 4.2 to define q_0.5; this is a standard threshold in structural biology but its application to astronomical images is a domain assumption.
  • domain assumption The 63.1 deg2 Q1 overlap region is representative of the full Euclid DR1 footprint in terms of galaxy populations and imaging properties.
    Implicit in the release of E-BGS over the full DR1 footprint (Section 5); not explicitly tested.
  • domain assumption Single-component Sersic profiles with chi2_nu < 3 adequately describe the clean subset of galaxies for parameter recovery.
    Used in Section 4.3 to define the fitting subset; standard but excludes mergers and irregulars.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From DESI to Euclid: A Generative Bridge to Improve Measurements of Galaxy Structure." pith.science (2026). https://pith.science/paper/V5HQLKBS

@misc{pith2026260706891,
  author       = {Pith},
  title        = {Pith review of: From DESI to Euclid: A Generative Bridge to Improve Measurements of Galaxy Structure},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5HQLKBS}},
  note         = {Machine review of arXiv:2607.06891}
}
abstract

Ground-based seeing imprints size-dependent biases on galaxy structural parameters, yet the high-resolution space-based imaging needed to improve these measurements currently covers only a small fraction of the sky. We close this gap with a generative model that produces Euclid-like VIS predictions from DESI imaging of Bright Galaxy Survey (BGS) targets. Our predictions remain Fourier-correlated with Euclid VIS images down to 0.37'', compared with 1.41'' and 1.00'' for the DESI $r$- and $z$-band inputs, corresponding to improvements by factors of ${\sim}3.8$ and ${\sim}2.7$, respectively. Although this correlation does not extend down to 0.16'', the characteristic Euclid VIS PSF FWHM, structural measurements from these predictions already show reduced biases relative to the DESI $r$-band structure measurements: the Petrosian radius bias falls to +0.072'' (from -0.845''), independent of galaxy size; the bias in the S\'ersic effective radius ($R_{\rm e}$) drops to -0.018'' (from -0.322''); and the S\'ersic-index bias to +0.093 (from +0.262). We release these predictions over the Euclid DR1 footprint as the Euclid-like Predictions of BGS (\textbf{E-BGS}), which can be blindly validated once DR1 is public.

Figures

Figures reproduced from arXiv: 2607.06891 by the authors.

Figure 1
Figure 1. Transmission curves for the four DECam bands (g, r, i, z; representative of the Legacy Surveys filters) and the Euclid VIS channel (∼550–900 nm). 2.2. Euclid The Euclid mission (R. Laureijs et al. 2011; E. Collab￾oration et al. 2024) performs a wide-area survey with the VIS imager (E. Collaboration et al. 2025a) and the Near￾Infrared Spectrometer and Photometer (NISP). Quick Data Release 1 (Q1; E. Collaboration et a… view at source ↗
Figure 2
Figure 2. The Image-to-Image Schr¨odinger Bridge (I2SB) for DESI→Euclid translation. The image sequence runs between two fixed endpoints, the DESI r-band source x1 (left) and the Euclid VIS target x0 (right), through intermediate states xt. The blue arrow marks the forward bridge interpolation (Eq. 1); the orange arrow marks the learned reverse sampling, in which the UNet ϵθ, conditioned on the DESI z-band (c = xz), maps xt t… view at source ↗
Figure 3
Figure 3. Predictions for randomly selected test-set galaxies (labeled by R.A., Dec.). Columns: DESI r-band, DESI z-band, Euclid Q1, prediction, and residual (prediction − Euclid) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Frequency-domain recovery against the GT over the test set (median; shaded 16–84th-percentile bands). (a) Azimuthally averaged power spectrum P versus an￾gular scale for the GT, E-BGS, and the DESI r and z bands. (b) Fourier Ring Correlation with the GT for E-BGS and t…
Figure 5
Figure 5. Figure 5: Structural parameters versus the GT, one per row: elliptical Petrosian radius RPet (top), S´ersic effective radius Re (middle), and S´ersic index n (bottom). Columns compare E-BGS (left), DESI r-band (center), and DESI z-band (right) against the GT; points are color-co…
Figure 6
Figure 6. Figure 6: Randomly selected E-BGS predictions over the Euclid DR1 footprint, where no Euclid imaging yet exists, so no ground truth is available. Each pair of columns shows the DESI grz composite (used for display only; the model takes the r and z bands) and the corresponding E-…

Discussion (0). Continue with ORCID to comment.

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.