{"id":"a3d83ff2-779e-44b6-a8cc-f3d9099bcd6f","arxiv_id":"2411.18440","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A redshift-conditioned diffusion model generates galaxy images whose ellipticity, size, Sersic index, and isophotal area track the morphology-redshift trends observed in the Hyper Suprime-Cam survey.","lead":"This paper trains a diffusion model to generate galaxy images conditioned on a continuous redshift value, and shows the generated galaxies reproduce known trends in shape and size with distance. It is a technical demonstration that morphology encodes redshift, but it does not actually measure redshifts from real images.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The morphology-z evidence in Figs. 3/4 depends on an undocumented metric pipeline: no PSF, threshold, deblending, or fitting details are given for ellipticity, semi-major axis, Sersic index, or isophotal area, so the match may reflect the estimator rather than the learned conditional distribution.","rationale":"I read the paper as claiming a continuous redshift-conditioned generative model whose samples reproduce known morphology-redshift trends, and as using that reproduction as evidence for an implicit understanding of galaxy evolution. The architecture and training are described in enough detail to be plausible; the dataset is public and the split is specified; the inclusion of a comparison to binned or discrete conditioning is a useful ablation. The weakest link is not the diffusion machinery but the evaluation metrics. Because the paper never describes the measurement software or parameters, I cannot tell whether the close match in Figs. 3/4 reflects physical fidelity or a shared estimation bias. The authors do flag high-z sparsity and environmental effects as limitations, but they do not flag this measurement comparability problem, and it is more fundamental: it concerns the validity of the main quantitative result at all redshifts. The reader's weakest assumption identifies the same issue, and I agree with the conditional verdict: the manuscript should be accepted only if the metric pipeline is specified and the comparison is redone with a fixed, public implementation. This is not an internal inconsistency but a missing-support and omitted-details concern; the proposed control experiment would settle it.","tokens_in":12531,"tokens_out":4185,"duration_ms":41864,"concrete_test":"Recompute all four metrics on both real HSC test cutouts and DDPM-generated images using one public, fixed pipeline (e.g., statmorph or SExtractor plus GALFIT) with explicitly stated parameters, including PSF, detection threshold, and band choice; then compare per-redshift-bin means and full distributions. As a control, run the same pipeline on real images smoothed or convolved to match the effective PSF and noise level of generated images; if the control reproduces the generated-vs-real differences, the claimed morphology-z match is a measurement artifact rather than evidence about the conditional distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the DDPM learns p(X|z) and reproduces morphology-z trends; the quantitative support is the close match of ellipticity, semi-major axis, Sersic index, and isophotal area between generated and real test images (Figs. 3/4, Table 1). Section 4 says only 'We compute standard morphological metrics' and lists textbook definitions; it never states the software, the detection threshold for isophotal area, the background-subtraction method, the PSF model or convolution for Sersic fitting, the deblending treatment, or whether metrics are measured on a single band or the five-band stack. Real HSC cutouts contain sky noise, correlated pixel noise, and a wavelength-dependent PSF; DDPM samples are produced by denoising from Gaussian noise and have different noise and background statistics. Ellipticity and isophotal area are threshold-sensitive, and Sersic index is strongly PSF-biased unless the PSF is modeled. Thus the agreement in Figs. 3/4 could be produced by the measurement pipeline (e.g., thresholding real noise and smooth generated backgrounds in the same way) rather than by an accurate learned morphology-z relation. This is the load-bearing assumption because every physical claim and the redshift-estimation application rest on those metrics.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a denoising diffusion probabilistic model (DDPM) trained on 64x64 five-band Hyper Suprime-Cam galaxy cutouts, conditioned on a log-transformed continuous redshift, with Gaussian noise added to the conditioning value during training. The authors generate galaxies at test-set redshifts and compare them to real test galaxies using four morphological metrics: ellipticity, semi-major axis, Sérsic index, and isophotal area. They report close agreement of the metric distributions (Figure 3) and of mean trends versus redshift (Figure 4), and a strong correlation between a CNN-predicted redshift and the conditioning redshift (Figure 5). The paper additionally benchmarks against discrete- and continuous-conditioned GAN and DDPM baselines (Table 1) and presents visual comparisons. It claims to establish a 'first direct link between galaxy morphology and redshift' and to offer a path to redshift estimation from imaging data.","tokens_in":12810,"tokens_out":7785,"duration_ms":67303,"significance":"If the morphological agreement shown in Figures 3 and 4 holds under a properly specified measurement pipeline, the paper provides a useful demonstration that a redshift-conditioned generative model can learn morphology-redshift correlations from imaging alone, which is potentially valuable for simulation and survey validation. The continuous conditioning scheme with perturbation is a sensible technical contribution, and the ablation over the perturbation scale sigma (Table 1) is informative. However, the paper's broader claims (a 'first direct link' and a 'simple and effective path to redshift estimation') are not supported by the evidence, and the lack of a documented morphological measurement pipeline is a serious gap. The strengths are the public code release, the benchmark comparisons, and the explicit treatment of continuous conditioning.","major_comments":[{"comment":"The central claim that the model reproduces morphology-z trends rests on the quantitative comparison in Figures 3 and 4, but the paper does not state how the four metrics are measured on either real or generated images. Ellipticity and isophotal area are threshold-dependent, the Sérsic index is strongly PSF-biased unless the PSF is modeled and convolved, and real HSC cutouts have sky noise, correlated pixel noise, and a wavelength-dependent PSF, whereas DDPM outputs are denoised from Gaussian noise and have different background statistics. Without specifying the detection threshold, background subtraction, deblending, PSF treatment, and band(s) used, the close match could be produced by the measurement pipeline rather than by the learned conditional distribution. Please provide a complete description of the metric pipeline (software and parameters), or adopt a public pipeline such as SExtractor/GALFIT/Statmorph, and validate the measurement on simulated images of known input parameters.","section":"Section 4 (morphological metrics)"},{"comment":"The redshift-prediction evidence in Figure 5 is partially circular and does not support the claim of a 'simple and effective path to redshift estimation'. Generated images are conditioned on the redshift z, so a predictor trained on real images will correlate with the conditioning value whenever the model has learned any z-dependent signal (e.g., overall color or flux), not necessarily morphology. The predictor CNN is from the authors' own prior work and may have been trained on the same data set, and the paper gives no comparison with photometric redshift baselines, no bias/scatter statistics, and no test on a completely independent sample. In addition, the abstract's phrase 'first study to establish a direct link between galaxy morphology and redshift' ignores the substantial existing literature on morphology-redshift relations (e.g., Conselice 2014); the correct claim is that a conditionally trained DDPM reproduces known trends.","section":"Abstract and Section 4 (Figure 5)"},{"comment":"Table 1's physical benchmark columns are ratios of the average synthetic metric to the average real metric. A degenerate model that always outputs an image with the average ellipticity, size, Sérsic index, and area would score near 1 on all four columns while failing to match the joint distribution. The paper should supplement these global-mean ratios with per-redshift-bin distribution comparisons (e.g., Kolmogorov-Smirnov or energy distances) and report per-bin mean ratios with confidence intervals. This is necessary to support the claim of 'notably better performance in terms of the physical benchmarks' in Section 4.1.","section":"Table 1"},{"comment":"Sections 2 and 5 state that 92.8% of the sample has z < 1.5 and that the model's performance is 'notably less reliable at higher redshifts', yet the title and abstract claim insight 'through cosmic time'. The redshift bins in Figure 4 should be shown with per-bin sample sizes and confidence intervals, and the high-z bins (z > 1.5) should either be demonstrated to reproduce the trends or explicitly excluded from the claim. Without this, the 'through cosmic time' claim is stronger than the evidence supports.","section":"Sections 2 and 5; Figures 3/4"}],"minor_comments":[{"comment":"The abstract contains grammatical issues: 'learns this missing joint structure, reproduces known morphology-z correlations' is missing 'and' before 'reproduces', and 'we verify on the HyperSuprime-Cam survey, that the model captures' contains a comma splice.","section":"Abstract"},{"comment":"Typos and incomplete phrasing appear: 'knowldedge', 'appoaches', and 'we explore the model's ability under to generate galaxies conditioned at different cosmic time-points' should be corrected.","section":"Section 1"},{"comment":"'prohibitively compute and time-expensive' should be 'prohibitively computationally and temporally expensive'.","section":"Section 2"},{"comment":"There are several typos and grammatical errors: 'bechmark', 'HyperCam-Suprime', and 'Our model achieve the second best FID score our of the model that were compared' need revision.","section":"Section 4.1"},{"comment":"'V AE's' and 'VQ-V AE's' should be 'VAEs' and 'VQ-VAEs', and 'Moreso' should be 'Moreover'.","section":"Section 5"},{"comment":"The reference 'Lizarraga et al. 2024a' cited in the Introduction appears to be this same manuscript (arXiv:2411.18440); please replace it with the intended prior work or mark it as a self-citation appropriately.","section":"References"},{"comment":"The cautionary paragraph about RGB projection and dust/heat artifacts is placed in the caption but would be clearer in the main text, and the sentence structures should be polished.","section":"Figure 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's central technical idea is plausible, but the authors overstate its novelty and the redshift-estimation implication. The 'first direct link' claim is inaccurate in the context of the astronomical literature, and the paper does not compare with photometric redshift methods. The self-citation of the same arXiv preprint as prior work should be checked. For this journal, I would suggest requiring the morphological pipeline description and a substantial toning-down of the claims before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper trains a DDPM conditioned on continuous redshift with Gaussian label perturbation and evaluates it on HSC galaxies. The genuinely new piece is the continuous conditioning trick and the evaluation protocol that compares physical metrics (ellipticity, semi-major axis, Sersic index, isophotal area) of generated vs real images, including ablations against discrete conditioning. That part is useful: Table 1 shows continuous conditioning with sigma=0.1 yields physical-metric ratios close to 1 and beats both discrete conditioning and GAN baselines. The low-redshift morphology trends reproduce known results, which is a reasonable sanity check. Credit where due: the authors put the code up and acknowledge the high-redshift sparsity and that FID is dubious for multiband data.\n\nThe soft spots are real. The biggest one is the unspecified measurement pipeline. The paper says \"We compute standard morphological metrics\" and gives textbook definitions, but never states what software was used, how the PSF was handled, what threshold defined the isophotal area, how background subtraction or deblending worked, or whether metrics came from a single band or a stack. Sersic indices are strongly PSF-biased unless the PSF is modeled; isophotal areas are threshold-sensitive. Since DDPM samples have different noise and background statistics than real cutouts, the agreement in Figures 3 and 4 could partly reflect the estimator, not the learned conditional distribution. That undercuts the load-bearing evidence.\n\nThe \"first direct link between galaxy morphology and redshift\" overclaims; morphology-redshift correlations are established, the model reproduces them. The redshift-prediction application also isn't validated on real images: the predictor comes from the authors' own earlier work, and generated images are conditioned on the very redshifts the predictor recovers, so the strong correlation in Figure 5 is partly built in. A test on real data with predicted z is missing.\n\nThe prose also leans on a large number of self-citations, which is noisy but not a substantive flaw.\n\nNet: the artifact is new and useful for survey simulation, but the paper as written does not nail the physical claim because the metric pipeline is undocumented. It deserves a serious referee, but I'd send it back for major revision: document the measurement code, validate the pipeline on noised or simulated images, drop the \"first direct link\" language, and either remove the redshift-prediction claim or validate it on held-out real galaxies. I wouldn't cite this version in my own work until those details are in.","headline":"A useful continuous-conditioning DDPM for galaxy simulation, but the physical-match claim rests on an undocumented metric pipeline and overclaimed novelty.","tokens_in":13332,"tokens_out":2436,"would_cite":false,"duration_ms":23501,"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":"A redshift-conditioned diffusion model learns galaxy morphology trends from images alone.","keywords":["galaxy morphology","redshift","diffusion models","generative models","photometric redshift","galaxy evolution","Hyper Suprime-Cam","Sérsic index"],"falsifier":"Recompute all four morphological metrics on both real and generated images with a single, publicly documented pipeline using identical background subtraction, thresholding, PSF convolution, and deblending; if the redshift-binned means of the generated images then depart from the real means beyond the 95% confidence intervals shown in Figure 4, the paper's claim of physical fidelity is falsified. A second check: train a CNN redshift estimator on real images only, then verify that on generated images its predictions correlate with the conditioning redshift after matching the redshift distribution; a null result would directly contradict the claim that generated images carry a redshift signal.","tokens_in":12340,"feed_emoji":"🌌","tokens_out":10024,"duration_ms":71917,"temperature":0.7,"pith_summary":"This paper claims that a denoising diffusion model conditioned on a continuous redshift value can learn the joint structure linking a galaxy's distance to its appearance, and that the images it generates reproduce known morphology–redshift correlations. The authors train on Hyper Suprime-Cam galaxy images with spectroscopic redshifts and show that, across redshift bins, the mean ellipticity, semi-major axis, Sérsic index, and isophotal area of generated galaxies closely track the real test distribution. They also show that a CNN redshift predictor trained on real galaxies assigns redshifts to generated images that correlate with the conditioning redshift, indicating a genuine redshift signal in the output. If correct, this offers a path to photometric redshift estimation that uses morphological information, and a way to simulate physically plausible galaxy populations at arbitrary redshifts.","feed_headline":"Diffusion model learns galaxy shapes from redshift alone","feed_subtitle":"A continuous conditioning approach lets generated galaxies match real ellipticity, size, and Sérsic trends.","key_machinery":"The key mechanism is a redshift-conditioned denoising diffusion probabilistic model (DDPM) built on a U-Net, in which the conditioning variable is a continuous, log-transformed redshift z' = log(1+z) embedded via sinusoidal positional encoding along with the diffusion timestep. During training, Gaussian noise z' + N(0,σ) is added to the redshift value, forcing the model to interpolate smoothly across nearby redshifts rather than memorize discrete bins. This perturbation lets the model learn p(X|z) over a continuum, and the paper shows that a small perturbation (σ=0.1) yields physically accurate morphological metrics, whereas σ=0 (discrete conditioning) sharply degrades them.","core_discovery":"Continuous redshift conditioning enables a denoising diffusion model to learn the conditional distribution of galaxy images given distance, and the generated images reproduce the redshift-dependent trends in ellipticity, semi-major axis, Sérsic index, and isophotal area observed in real Hyper Suprime-Cam data. The authors train a U-Net DDPM on 64x64 five-band galaxy images with log-transformed redshift z' = log(1+z) as a continuous condition, adding Gaussian noise of standard deviation σ to the redshift during training. With σ=0.1 the model's generated images have morphological metric ratios of 0.93–0.98 relative to the real test-set means, outperforming discrete-conditioned baselines and continuous GANs on these physical metrics, despite a slightly worse FID than the discrete DDPM baseline. The paper interprets this as evidence that the model implicitly encodes physical galaxy structure and its evolution from photometry and redshift alone, without morphological labels.","pith_inferences":["A direct extension the paper does not pursue is to use the learned conditional model itself as a photometric-redshift estimator by searching for the conditioning redshift whose generated image matches a target image; the reported correlation between conditioning and CNN-predicted redshift suggests this inversion could be competitive, but the paper does not test it.","The smoothness of the learned conditional distribution could be quantified by generating a finely spaced redshift trajectory from a single seed and measuring how much the morphological metrics change between consecutive redshifts; the paper asserts smooth transitions but does not measure their continuity.","The comparison of real and generated morphological metrics would be more convincing if the measurement software were specified and applied identically to both sets; because that is absent, a follow-up using a fixed catalog-level pipeline would separate physical fidelity from pipeline consistency.","The model's increased scatter at high redshift, which the paper attributes to sparse training data, suggests a testable prediction: conditioning on ancillary physical variables (stellar mass, environment) alongside redshift should reduce that scatter if morphology evolution is driven by more than redshift alone."],"forward_implications":["If correct, morphologically informed photometric redshift estimation becomes possible by inverting or conditioning such a generative model, potentially breaking degeneracies that pure color-based templates miss.","Astronomers could generate synthetic galaxy populations at any requested redshift for survey simulations, including redshifts where spectroscopic training data is scarce, without needing to tune physical parameters by hand.","The continuous conditioning scheme suggests that future galaxy generative models should treat physical parameters like redshift as continuous variables rather than discrete bins, since discretization degrades the physical fidelity of the output.","The benchmark table implies that FID is a poor standalone metric for scientific galaxy generation; evaluating physical morphology is necessary, and the paper offers a template for doing so."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model framework (forward noising and learned reverse denoising) that the paper adapts for continuous redshift conditioning.","marker":"Ho et al. [2020a]"},{"why":"Provides the Hyper Suprime-Cam galaxy dataset with spectroscopic redshifts and the train/test split source for training and evaluation.","marker":"Do et al. [2024b]"},{"why":"Gives the CNN redshift predictor used to validate that generated images carry a redshift signal, and suggests the (g,r,i) subsampling for FID computation.","marker":"Li et al. [2024b]"},{"why":"Defines the training/test split of the HSC dataset that the paper follows (204,513 training and 40,914 test images).","marker":"Li et al. [2024a]"},{"why":"Serves as the discrete-conditioned DDPM baseline that the paper re-implements and compares against in its benchmarks.","marker":"Smith et al. [2022]"},{"why":"Serves as the GAN baseline that the paper re-implements with redshift conditioning and compares against on FID and physical metrics.","marker":"Margalef-Bentabol et al. [2020]"},{"why":"Provides the redshift loss |ẑ-z|/(1+z) used to evaluate redshift prediction quality on generated images.","marker":"Nishizawa et al. [2020]"},{"why":"Defines the morphological metrics (ellipticity, semi-major axis, Sérsic index, isophotal area) used to compare real and generated galaxies.","marker":"Conselice, 2014a"}],"fun_headline_variants":["Redshift alone shapes galaxies in diffusion model","Diffusion model learns galaxy morphology from redshift","Continuous redshift conditions diffusion for galaxy images","Galaxy structure reproduced via redshift-conditioned diffusion","Redshift drives galaxy shape generation in diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The morphological metrics computed on generated 64x64 five-band images are directly comparable to the same metrics on real HSC images, even though the paper does not describe the measurement software, background subtraction, thresholding, deblending, or PSF treatment used to derive ellipticity, semi-major axis, Sérsic index, and isophotal area.","fun_headline_variants_meta":{"raw":{"variants":["Redshift alone shapes galaxies in diffusion model","Diffusion model learns galaxy morphology from redshift","Continuous redshift conditions diffusion for galaxy images","Galaxy structure reproduced via redshift-conditioned diffusion","Redshift drives galaxy shape generation in diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000374,"raw_usage":{"total_tokens":1987,"prompt_tokens":926,"completion_tokens":1061,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":995}},"tokens_in":542,"tokens_out":1061,"duration_ms":9732,"temperature":1.0,"reasoning_tokens":995,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:12:21.182143+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute all four morphological metrics on both real and generated images with a single, publicly documented pipeline using identical background subtraction, thresholding, PSF convolution, and deblending; if the redshift-binned means of the generated images then depart from the real means beyond the 95% confidence intervals shown in Figure 4, the paper's claim of physical fidelity is falsified. A second check: train a CNN redshift estimator on real images only, then verify that on generated images its predictions correlate with the conditioning redshift after matching the redshift distribution; a null result would directly contradict the claim that generated images carry a redshift signal.","supporting_citations":[{"cited_title":"Detecting outliers in astronomical images with deep generative networks","cited_arxiv_id":null,"evidence_quote":"Serves as the GAN baseline that the paper re-implements with redshift conditioning and compares against on FID and physical metrics."}],"review_version":1}