Pith. sign in

REVIEW 3 major objections 6 minor 6 references

Deblending Overlapping Galaxies in DECaLS Using Transformer-Based Algorithm: A Method Combining Multiple Bands and Data Types

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Deep-learning deblender hits 1 percent flux error on galaxy blends

desk verdict A solid synthetic evaluation of a known transformer architecture applied to galaxy deblending, but the headline real-data claims rest on a success criterion that does not verify flux or shape recovery. read the letter →

arxiv 2505.17452 v3 pith:DHIC7GHP submitted 2025-05-23 astro-ph.GA

Ran Zhang , Meng Liu , Zhenping Yi , Hao Yuan , Zechao Yang , Yude Bu , Xiaoming Kong , Chenglin Jia
show 3 more authors
Yuchen Bi Yusheng Zhang Nan Li
This is my paper · ORCID
classification astro-ph.GA
keywords galaxydeblendingDECaLStransformerU-Netdeeplearningimagerestorationmagnituderecoveryblending
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

Blending of overlapping galaxies is one of the main expected systematic uncertainties for upcoming deep surveys. This paper introduces CAT-deblender, a U-shaped transformer network trained on both RGB and grz-band images from DECaLS, which takes only approximate central coordinates of the target galaxy and reconstructs its flux and shape from the blended image. On artificial blends, the authors report a mean absolute percentage error of about 1 percent in recovering the central galaxy's magnitude for quadruply blended systems, compared with 4.8 percent for SExtractor. They also report deblending 433 real overlapping galaxy pairs and 63,733 of 100,000 randomly selected blended DECaLS images. If these results hold, the method offers a data-driven alternative to profile-model deblenders for real survey data.

What carries the argument

The central object is the Cross-Aggregation Transformer Block (CATB), which combines rectangular-window self-attention (computed in horizontal and vertical windows in parallel), an axial shift operation that widens the attention field, and a local complementary convolution that adds CNN-style locality. These blocks are arranged in a U-shaped encoder-decoder with skip connections; the model is trained with MSE loss on RGB images and Charbonnier loss on multi-band images, using transfer learning from the RGB-trained weights. The data engine is key: artificially blended images constructed from 2-4 superimposed isolated galaxies sampled from the GZD-5 morphological catalog, with randomized rotations, shifts, and scalings, so the network learns to reconstruct the central galaxy without needing the number or positions of neighboring sources.

What would settle it

Apply the trained CAT-deblender to a subset of the 100,000 real blends and compare the recovered central-galaxy flux against independent photometry of the same fields from space-based imaging; if a large fraction of cases show recovered magnitudes differing by more than the claimed ~1 percent while still passing the SExtractor single-detection check, the effective-deblending claim is falsified.

Watch

Extended reading notes

Core claim

The paper claims that a transformer-based generative network, CAT-deblender, can separate overlapping galaxies in DECaLS images using a U-shaped Cross-Aggregation Transformer architecture that requires only the approximate central coordinates of the target galaxy. The network is trained on artificially blended images built from isolated GZD-5 galaxies, in both RGB (Lupton-composed) and grz multi-band formats, and it learns to map a blended scene to a deblended central galaxy. In the test set, the recovered r-band magnitude has median error within ±0.25 and ellipticity within ±0.025, with pixel residuals below 0.005; magnitude recovery error stays at roughly 1 percent even when four galaxies are blended, versus 4.8 percent for SExtractor. On real data, the authors report successful deblending of 433 overlapping galaxy pairs cross-matched from SDSS and GZD-5 catalogs and of 63,733 randomly selected blended images from DECaLS, where success was judged by SExtractor recognizing a solitary galaxy after deblending.

Load-bearing premise

The claim that 63,733 real blends were effectively deblended rests on the assumption that a post-deblend image in which SExtractor detects only a single galaxy counts as a successful deblend, a condition that can also be met by removing the neighbor's flux instead of accurately reconstructing the central galaxy.

Editorial extensions

If this is right

  • CAT-deblender can be applied to survey images using only detection-stage central coordinates, sidestepping explicit modeling of the number or positions of neighboring sources.
  • Recovered r-band magnitudes and ellipticities of central galaxies have median errors within ±0.25 and ±0.025 respectively on the test set, with residuals below the typical r-band pixel threshold.
  • Unlike SExtractor, whose magnitude recovery error grows with blend multiplicity, CAT-deblender keeps its mean absolute percentage error near 1 percent for blends of up to four galaxies.
  • The method produced a catalog of deblended overlapping galaxies from DECaLS, including 433 pairs cross-matched with SDSS and 63,733 random blends, which could support photometric and morphological studies of blended systems.
  • If applied to future surveys, the method's positive bias in magnitudes and ellipticities would need correction before use in weak lensing and photo-z analyses.

Reading between the lines

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

  • The real-data success criterion (SExtractor detects only a solitary galaxy) likely overstates accuracy: it can be met by suppressing the neighbor's flux rather than faithfully recovering the central galaxy. The paper itself calls for more stringent assessment methods.
  • Because the training set comes from GZD-5, which is predominantly low-redshift, the model's performance on higher-redshift blends, where morphology differs, remains untested; transfer learning to other surveys would need redshift-dependent validation.
  • The reported positive bias in ellipticity and r-band magnitude, if uncorrected, would translate into a multiplicative shear calibration bias that is an order of magnitude above the ~0.001 requirement for LSST, so the method is not survey-ready without bias calibration.
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

3 major / 6 minor

Summary. The paper introduces CAT-deblender, a U-Net-structured transformer for deblending overlapping galaxies in DECaLS images, trained on synthetic blends of two to four galaxies created from isolated GZD-5 galaxies. The model is evaluated on RGB images with PSNR/SSIM and on grz multi-band images with ellipticity and magnitude recovery errors, and is compared against SExtractor and a GAN baseline. The authors also apply the model to real blended galaxies from DECaLS, reporting 433 successfully deblended cross-matched pairs and 63,733 effective deblends out of 100,000 randomly selected blends, with validation based on SExtractor detection and manual inspection.

Significance. If the central claims hold, the method would be a practically useful deblender for large surveys: it requires only approximate central coordinates, handles up to four blended sources, and shows lower magnitude recovery error than SExtractor on controlled synthetic tests. The paper's strengths include a substantial curated training set from GZD-5/DECaLS, explicit synthetic blend construction, a held-out test set, and a public catalog of real deblended systems. However, the real-data validation hinges on a detection-level criterion that does not verify flux or morphology recovery, the GAN baseline is not reproducible as reported, and the headline error rates are quoted without uncertainties. These gaps leave the real-data generalization claim under-supported, although the synthetic portion appears internally consistent.

major comments (3)
  1. [Section 4.3] The definition of 'effective deblending' as 'instances where the post-deblended images allowed SExtractor to recognize only a solitary galaxy' is insufficient to support the claim that 63,733 real blends were successfully deblended. This criterion can be satisfied by an output that removes the neighbor's non-overlapping flux, suppresses both sources below the detection threshold, or merges residuals into a single detection, without accurately recovering the central galaxy's flux or morphology. The paper's own Figure 13 shows 'successful but problematic' cases with missing overlap pixels and off-center peaks, and Section 5 concedes that 'Presently, we evaluate deblending effectiveness using tools like SExtractor and manual reviews. Future research should concentrate on stringent assessment methods.' Because the headline 1%-versus-4.8% magnitude comparison is measured on synthetic blends, the real-data extension of the central claim lacks photometric verification; please add an independent check (for example, comparing deblended fluxes and ellipticities with external measurements for a subsample, or at least quantifying the residual flux at the neighbor's position) before claiming successful deblending of the 63,733 and 433 real systems.
  2. [Table 2 and Section 4.2] Table 2 reports GAN baseline metrics (Mag_mape 1.79%, SSIM 0.96, PSNR 42.65) without specifying the implementation, training data, architecture details, or which checkpoint was used; as presented this comparison is not reproducible and its fairness cannot be assessed. In addition, all MAPE values in Table 2 and the abstract's '1 percent versus 4.8 percent' comparison are given without uncertainties, even though the test set is finite (12,000 multi-band images; 4,000 per blend multiplicity). The text is also inconsistent about which quantity the 1% figure refers to: the abstract says 'quadruply blended galaxies,' Section 4.2 says 'Even in the presence of three eccentric galaxies, CAT-deblender's MAPE remains a modest 1%,' while Table 2 lists a single overall CAT Mag_mape of 0.9%. Please report per-multiplicity MAPE values with bootstrap or other standard errors, and specify the exact GAN configuration used.
  3. [Section 2.1.1 and Section 4.2] The paper states that the training, validation, and test sets all have the same distribution and asserts in Section 4.2 that the model has not encountered the test galaxies, but it never describes how the split was constructed. If the split was not performed at the individual-galaxy level before synthetic blending, the same isolated galaxy image could appear in both training and test blends, which would make the reported PSNR, SSIM, and MAPE values optimistically biased. Please state explicitly whether the train/validation/test partition was done on galaxy IDs (or image cutouts) prior to creating the blended samples, and if so, describe the procedure.
minor comments (6)
  1. [Section 2.2] Equation (3) has a mismatched parenthesis in the displayed formula: it should be x_c = tanh(sinh^{-1}(x_raw,c / 255)). Please also define the constant epsilon in the Charbonnier loss in Equation (8).
  2. [Section 2.1.2] The description of source masking (threshold 2 sigma and connection zone of 3 pixels) is too terse; please specify whether these are SEP/Photutils parameters and whether the masks are dilated before being applied to remove 'unrelated astronomical entities.'
  3. [Section 4.1] The text says the model's PSNR/SSIM distributions have 'mean, median, apex, nadir, and variance,' but Figure 5 and the text only report the mean and standard deviation for PSNR and maximum/average SSIM; clarify what 'apex' and 'nadir' refer to.
  4. [Section 4.2] The sentence 'the peak signal-to-noise ratio's decrement markedly accentuates both the median error and its associated variance' is unclear; it should be rephrased to state what is being plotted in the lower panels of Figure 5.
  5. [General] There are frequent typographical issues, including 'refered' (Section 4.1), 'theCA TB' and 'theCA T-deblender' (Sections 3.2.1 and 3.2.2), and 'the peak position of the center galaxy is significantly off the center' (Section 4.3). The manuscript would benefit from a careful proofreading pass.
  6. [Section 4.3] The paper says the full deblending catalog is hosted on GitHub, but only a subset is shown in Table 3; please confirm that the complete catalog is publicly accessible at the stated link and provide the exact URL in the references.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported errors are held-out measurements, and the weak real-data validation criterion is a limitation rather than a circular reduction.

full rationale

The paper's central claims are empirical rather than derived from an analytic chain whose output is forced by its inputs. The synthetic test set is constructed from isolated galaxies that the paper states were not seen during training: 'CAT-deblender has neither encountered these particular combinations nor the possible individual galaxies.' Magnitude and ellipticity errors are obtained by comparing deblended outputs with pre-blend ground-truth images, and magnitude recovery is measured using SExtractor's sep.extract() on the deblended image rather than being directly optimized by the training loss, so the reported 1% versus 4.8% MAPE is a held-out measurement, not a fitted parameter renamed as a prediction. The Cross-Aggregation Transformer module is imported as an external architectural contribution from Chen et al. (2022), with no load-bearing self-citation or imported uniqueness theorem. The only concern is Section 4.3's 'effective deblending' definition, which requires only that SExtractor detect a solitary galaxy after deblending; this does not verify accurate flux or shape recovery on real data. That is a validation-protocol weakness, and the paper itself concedes in Section 5 that 'Presently, we evaluate deblending effectiveness using tools like SExtractor and manual reviews. Future research should concentrate on stringent assessment methods.' The criterion is not equivalent to the model's training objective and can fail, so the 63,733 count is an empirical measurement under a weak operational definition rather than a tautology or self-consistency result. No circular step is present.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard supervised learning assumptions, the linear blending model, and a specific curated training set. No new physical entity is introduced. The model weights are learned from data, not derived from first principles.

free parameters (3)
  • RGB interpolation scale constants = 0.04, 0.02, 0.1 in s = max(min(0.04p50, 0.02p90), 0.1)
    Eq. (1) sizes RGB cutouts; chosen by hand and affects how much galaxy light fits into the training frame, so it indirectly shapes what morphologies the model sees.
  • Photutils source detection threshold = 2 sigma with 3-pixel connectivity
    Section 2.1.2; determines which neighboring objects are masked out. A different threshold would change the training set and the effective isolation of 'isolated' galaxies.
  • GZD-5 morphological vote thresholds = see Table 1 (from Li et al. 2022)
    Used to select training galaxies by morphology; the distribution of morphologies in training depends on these thresholds.
assumptions (3)
  • domain assumption Linear blending model: observed flux is the sum of component galaxy fluxes along the line of sight.
    Stated in Section 1 and used throughout; galaxies are assumed not to interact and to be moderately opaque. If this fails, the network's target (additive superposition) is mis-specified.
  • domain assumption The GZD-5 morphological vote thresholds produce a representative sample of galaxy morphologies.
    Section 2.1.1, based on Li et al. (2022); the model's generality depends on the training set covering the morphology distribution of real blended galaxies.
  • ad hoc to paper Training, validation, and test splits do not share individual galaxy images.
    Section 4.2 states the model 'has neither encountered these particular combinations nor the possible individual galaxies'; this is asserted but not demonstrated with code or split metadata.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deblending Overlapping Galaxies in DECaLS Using Transformer-Based Algorithm: A Method Combining Multiple Bands and Data Types." pith.science (2026). https://pith.science/paper/DHIC7GHP

@misc{pith2026250517452,
  author       = {Pith},
  title        = {Pith review of: Deblending Overlapping Galaxies in DECaLS Using Transformer-Based Algorithm: A Method Combining Multiple Bands and Data Types},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DHIC7GHP}},
  note         = {Machine review of arXiv:2505.17452}
}
read the original abstract

In large-scale galaxy surveys, particularly deep ground-based photometric studies, galaxy blending is inevitable and poses a potential primary systematic uncertainty for upcoming surveys. Current deblenders predominantly rely on analytical modeling of galaxy profiles, facing limitations due to inflexible and imprecise models. We present a novel approach using a U-net structured transformer-based network for deblending astronomical images, which we term the CAT-deblender. It was trained using both RGB and grz-band images, spanning two distinct data formats from the Dark Energy Camera Legacy Survey (DECaLS) database, including galaxies with diverse morphologies. Our method requires only the approximate central coordinates of each target galaxy, bypassing assumptions on neighboring source counts. Post-deblending, our RGB images retain a high signal-to-noise peak, showing superior structural similarity to ground truth. For multi-band images, the ellipticity of central galaxies and median reconstruction error for the r-band consistently lie within +/-0.025 to +/-0.25, revealing minimal pixel residuals. In our comparison focused on flux recovery, our model shows a mere 1 percent error in magnitude recovery for quadruply blended galaxies, significantly outperforming SExtractor's higher error rate of 4.8 percent. By cross-matching with publicly accessible overlapping galaxy catalogs from the DECaLS database, we successfully deblended 433 overlapping galaxies. Furthermore, we demonstrated effective deblending of 63,733 blended galaxy images randomly selected from the DECaLS database.

Figures

Figures reproduced from arXiv: 2505.17452 by the authors.

Figure 1
Figure 1. The distribution of the number of galaxies in each class within the dataset, and the count of each galaxy morphology class in two data types: training set, validation set, and test set. The horizontal axis represents the types of galaxies, and the vertical axis represents the quantity of galaxy samples. classes in the dataset. The training set was used for model training, the validation set for adjusting the model’s… view at source ↗
Figure 2
Figure 2. Description of data preprocessing and manual mixing of samples: The (a) process illustrates the creation of masks, where three masks repre￾sent the grz-bands respectively. The (b) process showcases the masking of irrelevant objects. The mask represents the union of masks for irrelevant ob￾jects in the three-band. The third image displays the post-processed image that only contains the target central galaxy. The (c) … view at source ↗
Figure 3
Figure 3. The multi-head attention mechanism as defined in Vaswani et al. (2017) is described as follows. (a): The input sequence X is multiplied with the weight matrices WV , WK, and WQ, resulting in V, K, and Q respectively. (b): For multi-head attention, the value V, key K, and query Q each undergo h different learned linear projections. Subsequently, multi-head attention components generate h-dimensional output values in … view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: (a): Illustration of the rectangular window self-attention mechanism and axial shift operation. sh and sw are the window height and width of H-Rwin and V-Rwin, respectively. For a pixel (red dot), the attention area is the pink region in the middle feature map. (b): De…
Figure 5
Figure 5. Figure 5: elucidates, in its two lower subfigures, that in the r-band of the test dataset, the median error for ellipticity is consistently bounded within ±0.025, whereas the median er￾ror for the target galaxy magnitude does not exceed ±0.25. The impact of magnitude disparity b…
Figure 6
Figure 6. Figure 6: The examples of successful deblending and prediction on the test set for two data types are shown via CAT-deblender. The top three rows display examples of successfully deblended RGB images, while the bottom three rows show examples of successfully deblended r-band ima…
Figure 7
Figure 7. Figure 7: The first row of subfigures displays the magnitude error of central galaxies in blended galaxies measured using the direct deblending functionality of SExtractor, compared to their original magnitudes. The second row shows the magnitude error of central galaxies after …
Figure 8
Figure 8. Figure 8: The average residual pixel distribution of the r-band image unmix￾ing on the test set is shown. The top two subfigures display the unmixing average residuals for the bi-blended, tri-blended galaxies, while the bottom two subfigures display those for the quad-blended ga…
Figure 9
Figure 9. Figure 9: We present examples of RGB images of various overlapping galaxy types from the DECaLS database, along with the deblending results obtained using the CAT-deblender model. Each subplot consists of a pair of images. On the left side, we show the real observed blended gala…
Figure 10
Figure 10. Figure 10: We present examples of different mixed galaxy types in the r-band images from the DECaLS database, along with the deblending results obtained using the CAT-deblender model. Examples of successful unmixing in the r-band image from the multi-band images in the four pane…
Figure 11
Figure 11. Figure 11: We present a subset of successfully deblended data types selected from the DECaLS database. For each subplot, the top three panels represent the mixed galaxy image, the three-dimensional pixel value distribution of the mixed galaxy image, and the deblending mask obtai…
Figure 12
Figure 12. Figure 12: (a) A 128×128 three-band blended galaxy image sample from the DECaLS dataset, visualized through Lupton transformation (Lupton et al. 2004). (b): Activation on the first layer of the Encoder when a 3×128×128 image is input into the network. (c) Activation on the secon…
Figure 13
Figure 13. Figure 13: For those images that were successfully deblended but still have issues, we observed two distinct patterns: (i). There is a missing overlap area between the central galaxy and the eccentric galaxy in the deblended image, where the pixels are not appropriately filled, …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

6 extracted references · 2 canonical work pages

  1. [163]

    Hogg, David W, and Dustin Lang

    https://doi.org/10.3847/1538-3881/ab74cf . Hogg, David W, and Dustin Lang. 2013. Replacing standard galaxy profiles with mixtures of gaussians. Publications of the Astronomical Society of the Pacific 125 (928): 719. https://doi.org/10.1088/0067-0049/191/1/58. Huang, Song, Alexie Leauthaud, Ryoma Murata, James Bosch, Paul Price, Robert Lupton, Rachel Mande...

    work page doi:10.3847/1538-3881/ab74cf 2013

  2. [191]

    https://iopscience.iop.org/article/10.1086/131977. Stoughton, Chris, Robert H Lupton, Mariangela Bernardi, Michael R Blanton, Scott Burles, Francisco J Castander, AJ Connolly, Daniel J Eisenstein, Joshua A Frieman, GS Hennessy, et al. 2002. Sloan digital sky survey: early data release. The Astronomical Journal 123 (1): 485. https://doi.org/ 10.1086/324741...

    doi:10.1086/131977

  3. [241]

    https://link.springer.com/chapter/10.1007/978-3-319- 24574-4_28

    Springer. https://link.springer.com/chapter/10.1007/978-3-319- 24574-4_28. Samuroff, Simon, SL Bridle, Joe Zuntz, MA Troxel, D Gruen, RP Rollins, GM Bernstein, TF Eifler, EM Huff, T Kacprzak, et al. 2018. Dark energy survey year 1 results: the impact of galaxy neighbours on weak lensing cosmology with im3shape. Monthly Notices of the Royal Astronomical So...

    doi:10.1007/978-3-319- 2018

  4. [2018]

    IEEE transactions on pattern analysis and machine intelligence 41 (11): 2599–2613

    Fast and accurate image super-resolution with deep laplacian pyra- mid networks. IEEE transactions on pattern analysis and machine intelligence 41 (11): 2599–2613. https://doi.org/10.1109/TPAMI.2018.2865304. Larsson, Gustav, Michael Maire, and Gregory Shakhnarovich. 2017.Fractalnet: ultra-deep neural networks without residuals. arXiv: 1605.07648 [cs.CV]. ...

    arXiv 2018

  5. [2020]

    Multiband probabilistic cataloging: a joint fitting approach to point-source detection and deblending.The Astronomical Journal 159 (4):

    work page

  6. [2022]

    Physical Review D 106 (6): 063023

    Galaxy deblending using residual dense neural networks. Physical Review D 106 (6): 063023. https://doi.org/10.1103/PhysRevD.106. 063023. Wang, Zhou, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13 (4): 600–612. https://doi.org/10. ...

    arXiv 2004

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.