{"id":"bc6c49a3-efbc-444e-87b5-23536073a814","arxiv_id":"2501.15149","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A supervised CNN with L1 plus SSIM loss translates galaxy images across ultraviolet, optical, and infrared bands, with high validation-set similarity on Illustris mocks and moderate similarity on DECaLS data.","lead":"Using a convolutional image-to-image network, the authors show that galaxy images in one or two photometric bands can be converted into other bands, trained on simulated Illustris galaxies and tested on DECaLS survey data. The method is a proof of concept for filling in missing multi-band observations to help plan surveys and guide follow-up observations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The real-data result only shows same-survey G,Z→R interpolation; the UV/IR-wide claim rests entirely on Illustris mocks, so the Section 4 carryover assertion is an untested premise.","rationale":"I align with the reader's CONDITIONAL verdict. The strongest asset is a reproducible pipeline with public code, held-out validation sets, and morphology metrics (GINI/M20) that are external to the loss; the Illustris experiments are internally coherent and show large improvement over input-target SSIM. The load-bearing gap is not the method but the scope of the real-world claim. The DECaLS experiment is a valid proof of concept for same-survey optical interpolation, but it is the only real-data evidence and it does not exercise UV/IR translation or cross-instrument transfer. Section 4 (iii) overstates the evidence by asserting that Illustris successes carry over to observed data. A direct transfer experiment using G+Z→R is feasible with the public code and would settle whether the mock-to-real premise holds. Until then, the verdict should remain conditional, with the abstract and conclusions tempered to distinguish mock UV/IR fidelity from real optical interpolation.","tokens_in":12700,"tokens_out":5550,"duration_ms":56848,"concrete_test":"Train the same publicly available architecture on Illustris G and Z to predict R (the exact DECaLS band set) with matching asinh/normalization, then apply the Illustris-trained network directly to a held-out DECaLS test set without fine-tuning. Compare SSIM/PSNR/MAE against the DECaLS-trained model's Table 4 values (0.86 SSIM, 30.5 dB PSNR) and against the input-to-target baselines (0.53 and 0.38 SSIM). If the direct-transfer SSIM is close to 0.86, the carryover claim holds; if it falls much closer to the input baselines, Section 4 (iii) is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim has two linked parts: high-fidelity band translation, demonstrated only on Illustris mocks, and real-world predictive utility, demonstrated by one DECaLS model that interpolates R from G and Z inputs. The link is asserted in Section 4, contribution (iii), which states that Illustris successes 'carry over into observed datasets.' No experiment transfers an Illustris-trained model to real data; Section 3.3 trains a fresh model on DECaLS. Moreover, the DECaLS test uses closely spaced optical bands from a single survey, so it cannot validate translation involving ultraviolet or infrared bands under real noise, PSF, astrometric misalignment, or flux-calibration differences. Those are precisely the factors the mock data omit, and the paper's own future-work paragraph names 'noise variability and calibration differences across surveys' as open limitations. The headline claim about mapping across UV/visible/IR therefore depends on an unverified premise: that mock-to-mock mappings survive instrumental domain shifts. If that premise is false, the conclusion reduces to a narrow same-survey optical interpolation, not the general capability stated in the abstract. This is a missing-evidence problem, not a mathematical inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a supervised image-to-image CNN, based on a ResNet/CycleGAN-style generator, trained with an L1 plus inverse-SSIM loss, to translate galaxy images between photometric bands. Using paired Illustris mock observations, the authors train interpolation models (NUV, K to G, R, Z) and extrapolation models (G, R to U, NUV, FUV), reporting MAE, SSIM, PSNR, and Wasserstein distances between GINI/M20 distributions. A separate model is trained on DECaLS G and Z inputs to predict R. The paper claims high-fidelity band translation and potential for augmenting multi-band datasets, with code publicly available.","tokens_in":12863,"tokens_out":5820,"duration_ms":52897,"significance":"If the results hold, the paper is a useful proof of concept for using mock observations to prototype band-to-band translation and for augmenting sparse multi-band coverage. Strengths include held-out validation sets, the use of morphology metrics (GINI/M20) that are not part of the training loss, a clear architecture description, and public code. The main limitation is that the real-data demonstration is a narrow same-survey optical interpolation, so the broad UV/IR real-data claim in the abstract is not directly supported by the experiments.","major_comments":[{"comment":"The statement that Illustris successes \"carry over into observed datasets\" is not tested. Section 3.3 trains a fresh model on DECaLS G and Z inputs; no Illustris-trained model is applied to real observations. Moreover, the DECaLS experiment interpolates R from closely spaced optical bands within a single survey, so it does not exercise UV or IR translation under real noise, PSF, astrometric, or calibration conditions. The future-work paragraph acknowledges \"noise variability and calibration differences across surveys\" as open limitations. Please either add a transfer/fine-tuning experiment that directly tests cross-survey generalization, or temper the claim to separate mock-based feasibility from the narrow real-data interpolation.","section":"Section 4, contribution (iii); Abstract"},{"comment":"All reported metrics are point estimates without error bars, confidence intervals, or repeated runs. Given that training is stochastic and the validation sets are 500 (Illustris) and 2000 (DECaLS) images, differences such as SSIM 0.986 (Table 2, R) versus 0.99 (Table 3, U) may not be meaningful. Add uncertainty quantification (multiple seeds or bootstrap resampling) or at least report the spread over validation folds.","section":"Tables 2, 3, and 4"},{"comment":"The description of M20 as \"the relative contribution of the brightest 20 percent of a galaxy's light to its overall light distribution\" is inaccurate; M20 is the normalized second moment of the brightest 20% of the light, not a flux fraction. The paper also does not specify how GINI and M20 are computed (e.g., segmentation, aperture, pixel selection) or how the global normalization bounds for the [0,1] scaling were chosen. Since the W1 distances are the only morphology validation, these details are needed to assess whether the small values are meaningful.","section":"Section 2.2"},{"comment":"The GINI/M20 comparison is made only at the level of marginal distributions. Because the task is per-image translation, the paper should also report per-object GINI/M20 errors or a scatter/correlation plot between ground-truth and generated values; distribution-level agreement can hide systematic per-galaxy biases.","section":"Section 2.2 and Figures 2, 4, 6"}],"minor_comments":[{"comment":"The telescope name \"GALAX\" should be \"GALEX\" (Galaxy Evolution Explorer), both in the table caption and in the text.","section":"Table 1 and Section 2.1.1"},{"comment":"There are typographical errors: \"GINI coffecient\" should be \"Gini coefficient\" and \"earth movers distance\" should be \"Earth mover's distance\".","section":"Section 2.2"},{"comment":"The sentence \"As given in Table (2), the W1 distances...\" refers to DECaLS values that appear in Table 4, not Table 2.","section":"Section 3.3"},{"comment":"The column header \"λeff\" is never defined; clarify what effective lambda means and how it relates to λ in Eq. (2).","section":"Section 2.4 and Tables 2-4"},{"comment":"Specify the train/validation split procedure, including whether the split is random and whether galaxies are guaranteed not to appear in both training and validation sets.","section":"Appendix A.1"},{"comment":"The paper would benefit from a baseline comparison (e.g., linear interpolation between input bands or a simpler U-Net) to quantify the added value of the proposed architecture; currently the improvements over the input bands are shown, but not over a baseline model.","section":"Introduction and Results"}],"recommendation":"major_revision","confidential_remarks":"The paper is suitable in scope for an instrumentation/methods journal. The strongest issue is the gap between the abstract's general claim and the evidence: the broad UV/IR real-data claim rests on an untested carry-over assumption. I do not see a fatal internal inconsistency, but the load-bearing claim of transfer to real observations needs either new experiments or careful rewording. The absence of uncertainty estimates and the incomplete GINI/M20 methodological description should also be addressed before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper demonstrates a supervised image-to-image model that translates galaxy images between photometric bands on Illustris mocks with high SSIM/PSNR, and separately interpolates the R band from G/Z on DECaLS data with acceptable quality. That's a real and useful proof of concept, and it's honestly written. The GINI/M20 morphology metrics are external to the training loss, the code is public, and the writing is clear. Credit where due.\n\nThe soft spots are in the claim rather than the method. Section 4 lists contribution (iii) as 'successes with the Illustris dataset carry over into observed datasets,' but no experiment transfers an Illustris-trained model to real data. The DECaLS experiment trains a fresh model on G and Z from the same survey, and only interpolates a closely spaced optical band. That doesn't validate UV/IR translation under real noise, PSF, astrometric misalignment, or calibration differences—exactly the factors the authors themselves name as future-work limitations. So the abstract's promise of translating across UV, visible, and IR for real observations is supported only by mock-to-mock results plus a narrow same-survey interpolation. That's a missing-evidence problem, not a mathematical error.\n\nAlso missing: error bars or repeated runs, and baselines. MAE/SSIM improve a lot over the inputs, but there's no comparison against, say, a naive per-band scaling or a simple regression. Since the validation set was used for hyperparameter selection and then for the reported metrics, the numbers are likely optimistic. These are fixable in revision.\n\nBottom line: the core feasibility claim on mocks holds up, and the DECaLS result is a decent start. But the paper should be revised to either remove the carryover claim or add a real transfer experiment, include error bars, and benchmark against at least one simple baseline. I'd send it to peer review with the expectation of major revision. A reader in observational astronomy or survey planning will find it useful; a machine learning reader will find the methodology thin.","headline":"Solid mock-data proof of concept for galaxy band translation, but the real-data carryover claim is untested and the paper needs baselines and error bars before its broad conclusion is warranted.","tokens_in":13488,"tokens_out":2029,"would_cite":true,"duration_ms":18708,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a supervised generative image-to-image network can reliably translate galaxy observations between ultraviolet, visible, and infrared photometric bands, demonstrating the mapping on simulated Illustris galaxies and…","keywords":["Astronomy data modeling","Convolutional neural networks","Image-to-image translation","Galaxy photometry","Illustris simulations","DECaLS survey","Galaxy morphology","Generative deep learning"],"falsifier":"Take a model trained on Illustris mock pairs (for example, G and R inputs predicting NUV, or NUV and K inputs predicting R) and apply it without retraining to real galaxy images with matching coordinates, then compare the outputs against actual observations in the target band; if SSIM falls to roughly the level of the input bands' similarity to the target or the GINI/M20 distributions diverge substantially, the transfer claim collapses. A complementary test would be to train on DECaLS and extrapolate to a far-removed band such as FUV, comparing against existing GALEX imaging in the same fields.","tokens_in":12425,"feed_emoji":"🌌","tokens_out":6637,"duration_ms":52849,"temperature":0.7,"pith_summary":"This paper argues that a generative convolutional network can synthesize a galaxy image in one photometric band from images taken in other bands, effectively translating observations across ultraviolet, optical, and infrared light. The authors train a supervised ResNet-style generator on paired galaxy images from the Illustris simulations and report near-identical reconstructions for both interpolation (NUV and K inputs producing G, R, and Z bands) and extrapolation (G and R inputs producing U, NUV, and FUV bands). They also train the same architecture on real DECaLS images to recover the R band from G and Z band inputs, with somewhat lower but still high similarity scores. If the approach holds, astronomers could fill missing multi-band coverage, prioritize follow-up observations, and plan missions using synthesized images.","feed_headline":"Neural networks translate galaxy images across photometric bands","feed_subtitle":"Simulated galaxies are reconstructed with SSIM up to 0.99; real DECaLS data reach 0.86.","key_machinery":"The central object is a supervised generative convolutional network whose generator follows the CycleGAN generator design with a ResNet backbone, modified to use one downsampling and one upsampling block plus nine residual blocks to preserve high-frequency structure such as spiral arms and bars. It is trained with the compound loss $L = L_1 + \\lambda(1 - \\mathrm{SSIM})$, combining per-pixel mean absolute error with a structural similarity term that penalizes mismatches in local image structure. Paired, perfectly aligned Illustris mock images spanning FUV to K bands supply the training signal, while GINI and M20 morphology statistics, compared through the first Wasserstein distance, verify that generated images reproduce physical galaxy morphology rather than only low-level pixel statistics.","core_discovery":"A supervised image-to-image model with a ResNet-like generator and a combined L1 plus SSIM loss learns to map galaxy images between photometric bands, performing both interpolation and extrapolation with a single architecture that is retrained for each input-output band combination. On Illustris mock observations, the generated images match ground truth with MAE below 1.3%, SSIM between 0.91 and 0.99, PSNR between 28 and 40 dB, and GINI and M20 distributions close to the targets as measured by Wasserstein distance. On real DECaLS data, interpolating the R band from G and Z inputs yields SSIM 0.86, PSNR 30.5 dB, and similar morphological agreement. The authors interpret these results as a proof of concept that missing photometric bands of galaxies can be generated from partial observations.","pith_inferences":["The paper leaves untested whether an Illustris-trained model transfers directly to real galaxies; its real-data experiment retrains the model on DECaLS, so the claimed carry-over of simulation success to observations is an assertion rather than a demonstrated result.","Because Illustris pairs are noiseless and perfectly aligned, simulation metrics likely overstate the fidelity achievable on real data for wide-band translations; the DECaLS result uses closely spaced bands and does not yet test UV-to-IR extrapolation on observations.","A natural stronger test would be to predict ultraviolet bands from real optical survey images and compare against existing GALEX observations in matched fields, which would directly measure real-world extrapolation performance.","If the method scales to large photometric surveys, it could also serve as a cross-survey consistency check, since a failure to translate between bands would flag calibration or registration discrepancies between datasets."],"forward_implications":["Multi-band galaxy information can be synthesized from partial observations, potentially augmenting surveys that lack coverage in certain wavelength ranges.","Band extrapolation, such as predicting ultraviolet images from optical inputs, could extend existing datasets into wavelengths that were never observed for a given galaxy.","The same model architecture applies to any input-to-output band combination with only retraining, so the method is agnostic to whether the task is interpolation or extrapolation.","Real survey data can serve as training material for the model, with noise and background artifacts degrading fidelity only modestly compared with clean simulations.","Reliable band-to-band translation could support mission planning and guide high-resolution follow-up observations by predicting structures not resolved in available data."],"supporting_citations":[{"why":"Provides the paired Illustris mock observations spanning FUV to K bands that form the training and validation data for most experiments.","marker":"Torrey et al. 2015"},{"why":"Describes the Illustris cosmological simulation that produced the synthetic galaxies used in the mock catalog.","marker":"Vogelsberger et al. 2014"},{"why":"Supplies the CycleGAN generator architecture and default hyperparameters that the paper modifies for its band translation models.","marker":"Zhu et al. 2017"},{"why":"Introduces the residual blocks that form the backbone of the generator and support identity-friendly feature learning.","marker":"He et al. 2016"},{"why":"Establishes the supervised paired image-to-image translation paradigm (pix2pix) that this work follows.","marker":"Isola et al. 2017"},{"why":"Provides the Galaxy10 DECaLS dataset through astroNN, used for the real-data training experiment.","marker":"Leung & Bovy 2018"},{"why":"Documents the DECaLS survey whose real galaxy images are used as the proof-of-concept on observational data.","marker":"Dey et al. 2019"},{"why":"Defines the GINI and M20 morphology statistics used to validate that generated images match galaxy morphology.","marker":"Lotz et al. 2004"},{"why":"Provides the SSIM index formulation used both in the training loss and as a quantitative evaluation metric.","marker":"Brunet et al. 2011"}],"fun_headline_variants":["AI translates galaxy images across UV, visible, and IR","Generative model fills in missing galaxy bands from simulations","Deep learning predicts galaxy colors across photometric bands","Simulated galaxies train AI to complete real galaxy images","Neural network learns to map galaxy images between bands"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the idealized, noise-free Illustris mock images are representative enough of real survey conditions that the high-fidelity translations demonstrated on them carry over to real galaxies, but this transfer from mock to observed data is asserted rather than directly tested.","fun_headline_variants_meta":{"raw":{"variants":["AI translates galaxy images across UV, visible, and IR","Generative model fills in missing galaxy bands from simulations","Deep learning predicts galaxy colors across photometric bands","Simulated galaxies train AI to complete real galaxy images","Neural network learns to map galaxy images between bands"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001117,"raw_usage":{"total_tokens":4611,"prompt_tokens":869,"completion_tokens":3742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":3666}},"tokens_in":485,"tokens_out":3742,"duration_ms":21719,"temperature":1.0,"reasoning_tokens":3666,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:34:20.432083+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a model trained on Illustris mock pairs (for example, G and R inputs predicting NUV, or NUV and K inputs predicting R) and apply it without retraining to real galaxy images with matching coordinates, then compare the outputs against actual observations in the target band; if SSIM falls to roughly the level of the input bands' similarity to the target or the GINI/M20 distributions diverge substantially, the transfer claim collapses. A complementary test would be to train on DECaLS and extrapolate to a far-removed band such as FUV, comparing against existing GALEX imaging in the same fields.","supporting_citations":[{"cited_title":"F., Vogelsberger, M., et al","cited_arxiv_id":null,"evidence_quote":"Provides the paired Illustris mock observations spanning FUV to K bands that form the training and validation data for most experiments."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the CycleGAN generator architecture and default hyperparameters that the paper modifies for its band translation models."},{"cited_title":"M., Primack, J., & Madau, P","cited_arxiv_id":null,"evidence_quote":"Defines the GINI and M20 morphology statistics used to validate that generated images match galaxy morphology."},{"cited_title":"R., & Wang, Z","cited_arxiv_id":null,"evidence_quote":"Provides the SSIM index formulation used both in the training loss and as a quantitative evaluation metric."}],"review_version":1}