{"id":"327aabf6-e4a0-4b58-b4fb-7a13bc948034","arxiv_id":"2411.09356","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"WMGM generates 128x128 images by diffusing only low-frequency wavelet coefficients and using a shared multi-scale GAN to fill in high-frequency details, improving FID and cutting sampling time and parameters versus SGM and WSGM.","lead":"The paper combines two generative AI techniques in the wavelet domain: a diffusion model for coarse low-frequency image structure and a GAN for fine high-frequency details. It reports faster sampling and better image quality than two wavelet diffusion baselines on face, animal, and tissue images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's speed advantage rests on an unmeasured claim that the coarsest LL band is near-Gaussian and well-conditioned; the supporting CLT argument is not quantified at the actual scale, so the 16-step diffusion may still inherit the ill-conditioning it is meant to remove.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing point: the method's speed and quality advantage depends on the coarsest LL band being approximately Gaussian and well-conditioned, so that a short score-based diffusion captures global structure. My reading of the paper confirms this is the least secure part of the argument. Section 3.2.2 and Appendix E.1 give only an informal generalized-CLT justification and a qualitative whitening argument; they do not measure the condition number or Gaussianity of the LL coefficients at the scale and resolution actually used. Since Proposition 2, which is meant to support the high-frequency side, has an invalid proof, the theoretical scaffolding is weaker than the experimental section suggests. The empirical result may still be correct, but the central mechanism is not verified. A direct measurement of κ(Σ_LL) and an oracle-score LL-only sampling experiment would settle whether the 16-step diffusion is actually operating in the well-conditioned regime claimed. This does not require changing the reader's CONDITIONAL verdict, because the concern is already reflected in that verdict and in the requested additional validation.","tokens_in":18918,"tokens_out":6360,"duration_ms":68006,"concrete_test":"Instrument the released code to compute, for the exact S used in Table 1, the empirical covariance of the coarsest LL subband on each dataset after standardizing total energy Tr(Σ_LL)=d; report κ(Σ_LL) and the KL divergence between the best-fit Gaussian and the empirical LL distribution. Then replace the learned score on the LL band with the oracle score of that fitted Gaussian and run the same 16-step reverse process: if the oracle LL samples fail to preserve global structure, or if κ(Σ_LL) is comparable to the full-image condition number, the 'well-conditioned low-frequency' premise is not supported. A weaker follow-up is to train WMGM with S=2, 3, and 4 and check whether the FID at 16 steps degrades as S shrinks.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that moving diffusion to the wavelet domain removes ill-conditioning, so that a 16-step reverse process on the coarsest LL band suffices. This requires the LL distribution to be approximately Gaussian with a well-conditioned covariance. The paper's only support is Section 3.2.2 and Appendix E.1, which invoke a generalized CLT on block averages. That argument is not made quantitative: it never specifies the actual coarsest scale S used in Table 1, nor does it report the LL covariance condition number or the KL divergence to a Gaussian for CelebA-HQ, AFHQ-Cat, or Colon. For 128x128 inputs at realistic S (e.g., 3 or 4), each LL coefficient averages only 8x8 or 16x16 correlated pixels; CLT convergence can be weak, and residual low-frequency power still leaves a wide eigenvalue spread. If kappa(Σ_LL) is not small, Eq. 9 (N <= ε^{-2}κ^3) still forces many steps, and the 16-step success cannot be attributed to the well-conditioning claim. The reported FID gains could then be coming from the multi-scale adversarial HF generator rather than from the wavelet-domain score component, which would decouple the experimental headline from the theoretical motivation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WMGM, a multi-scale generative model that operates in the wavelet domain. The method applies score-based diffusion only to the coarsest low-frequency (LL) wavelet band and uses a multi-scale adversarial network to generate high-frequency bands conditioned on the low-frequency information. The authors claim theoretical support for the well-conditioning of LL coefficients and for the sparsity/non-Gaussianity of high-frequency coefficients, and they report experimental results on CelebA-HQ, AFHQ-Cat, and Colon showing lower FID than SGM and WSGM at 16 total sampling steps, with 89M parameters and 2.31 seconds per 100 images. They also include an ablation study of multi-scale versus single-scale adversarial learning.","tokens_in":19230,"tokens_out":3843,"duration_ms":37645,"significance":"If the empirical claims hold, the paper offers a practically useful combination of diffusion and adversarial generation in the wavelet domain, with substantially fewer parameters and faster sampling than the compared baselines. The manuscript provides comparisons on three datasets, a promising anonymous code release, and an ablation of the proposed MSAL component. However, the theoretical contribution is partially inherited from prior work, one of the new proofs contains an invalid step, and the central well-conditioning premise is not directly measured at the scale used in the experiments. The significance is therefore conditional on strengthening the theoretical and empirical evidence.","major_comments":[{"comment":"The proof of Proposition 2 is invalid. From the average lower bound ∫ E(s(x_H^k)|x_L^k) p(x_L^k) dx_L^k ≥ α, the proof asserts that E(s(x_H^k)|x_L^k) has a uniform lower bound α' for all x_L^k. An average lower bound does not imply a pointwise lower bound, and no argument is given for such a uniform bound. Equation (70) then identifies α' with the integral minus ε, which is not justified. Consequently, Proposition 2, which is used to support the claim that conditional high-frequency distributions are non-Gaussian and sparse, is unproved as written. The authors should either provide a correct proof with explicit assumptions or restate the proposition as an empirical observation supported by Figure 6.","section":"Appendix E.2, Proposition 2"},{"comment":"The central theoretical motivation relies on the claim that the coarsest LL band is near-Gaussian and well-conditioned, so that a short 16-step reverse diffusion suffices. This claim is not validated at the actual scale used in the experiments. Appendix E.1 reports KL divergences only for scales 0, 1, and 2, and Figure 5 does not show the coarsest scale S used in Table 1. The paper also never reports the condition number κ(Σ_LL) or the KL divergence to a Gaussian at the operational S. Without these measurements, Eq. (9) cannot be used to explain the 16-step success, and the observed FID gains could be dominated by the multi-scale adversarial high-frequency generator rather than by the claimed well-conditioned diffusion component. Please report S and the relevant diagnostics for the datasets used in Table 1.","section":"Section 3.2.2 and Appendix E.1"},{"comment":"The central empirical comparison reports single FID values without error bars or repeated seeds. For a claim of significant improvement over SGM and WSGM, the authors should provide means and standard deviations over multiple training runs or at least over multiple evaluation batches. The sampling time is averaged over 10 trials, but FID is not, and the reader cannot assess the stability of the reported differences, especially for the smaller gaps such as 25.38 vs 26.74 on CelebA-HQ (30K).","section":"Table 1 and Figure 4"},{"comment":"The paper states that Theorem 1 and Proposition 1 are proved by referring to 'Proof S5 in [19]', and Theorem 6 is quoted directly from [19]. Since one of the stated contributions is to 'theoretically establish the generative modeling in the wavelet domain', the manuscript should either provide self-contained proofs for the statements it presents or clearly identify which results are prior work and what exactly is new in this paper. As written, the theoretical section does not substantiate the claimed new theoretical contribution beyond the earlier WSGM analysis.","section":"Appendix A and Section 3.1"}],"minor_comments":[{"comment":"The sentence 'our model significantly improve performance' contains a subject-verb agreement error and should read 'our model significantly improves performance'.","section":"Abstract"},{"comment":"The text refers to 'Appendix I' for the definition of MS and SS, but the definitions appear in Appendix H; please correct the cross-reference.","section":"Section 4.3"},{"comment":"References [18] and [19] are the same paper and appear to be duplicated; please merge them into a single reference.","section":"References"},{"comment":"The word 'hierachical' is misspelled and should be 'hierarchical'.","section":"Figure 2 caption"},{"comment":"The term 'ill-positionedness' should be 'ill-conditionedness' for consistency with the rest of the paper.","section":"Section 3.3"},{"comment":"The sentence 'Detailed model information can be found in ?? as well as the source code' contains a placeholder '??' that should be replaced with a specific appendix or section reference.","section":"Appendix H.2"},{"comment":"The limitation statement says the method 'sacrifices a little image quality', yet Table 1 reports lower FID than both baselines; please clarify whether 'image quality' here refers to a different comparison or a qualitative trade-off rather than the FID metric.","section":"Appendix J"}],"recommendation":"major_revision","confidential_remarks":"The manuscript shows a practically promising result, but the theoretical section needs substantial repair and the empirical claims need uncertainty quantification. The duplicate reference [18]/[19] and the placeholder in Appendix H.2 suggest the paper was not fully polished. If the authors address the proof issue in Proposition 2, measure or properly qualify the LL-band conditioning claim, and add repeated-seed FID results, the paper could be acceptable after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: WMGM is a genuine architectural combination—score-based diffusion on the coarsest LL band plus a parameter-shared multi-scale GAN for the HF bands—and the reported FID gains at 16 sampling steps are consistent across CelebA-HQ, AFHQ-Cat, and Colon. The parameter and time reductions are real. This is worth engaging with, not dismissing.\n\nWhat the paper does well: the multi-scale factorization in Eq. 17 is clean, the shared-parameter MSAL is a useful trick (multi-scale beats single-scale with half the parameters, Table 2), and the empirics are internally coherent. WSGM is a fair baseline, and the paper beats it on all reported settings. The code is available.\n\nThe soft spots are real but not fatal. The biggest one is the load-bearing assumption that the coarsest LL coefficients are near-Gaussian and well-conditioned. The evidence offered is KL divergence to a Gaussian at scales 1 and 2 on two datasets (Fig. 5), not at the actual scale S used in Table 1, and no condition number is reported. The generalized-CLT argument in Appendix E.1 is hand-wavy at the relevant scale: with 128x128 inputs and S=3 or 4, each LL coefficient averages only 8x8 or 16x16 correlated pixels. The stress-test concern lands. If the LL covariance is still ill-conditioned, Eq. 9 forces many steps, and the 16-step success would come from the GAN rather than the wavelet score component.\n\nSecond, Proposition 2's proof is wrong. Boundedness of E(s|xL) does not give a uniform lower bound, and the chain of inequalities in the proof doesn't establish the claimed α−ε. The proposition may be true, but the proof as written is invalid. Third, the core theory is borrowed from Guth et al.; Theorem 1 and Proposition 1 are deferred to their proof. That is okay if cited, but it means the paper's own theoretical contribution is mostly the empirical demonstration.\n\nOn empirics: no error bars, no repeated seeds, and only two baselines. The FID numbers are not close enough to be decisively better at 30K CelebA-HQ (25.38 vs 26.74), so variance matters. Missing modern fast-sampling baselines (e.g., consistency models, diffusion-GAN hybrids) weakens the claim of \"significantly improve performance\" but not the core result.\n\nVerdict: this deserves a serious referee. The architecture is new, the experiments are straightforward, and the flaws are fixable in revision. I would ask the authors to report the actual S, the LL condition number, error bars, and a corrected Proposition 2 before acceptance.","headline":"A credible hybrid of wavelet-domain score diffusion and a shared multi-scale GAN that beats its two baselines at 16 steps, but the theoretical support for the central well-conditioning claim is thinner than it looks.","tokens_in":19771,"tokens_out":2383,"would_cite":false,"duration_ms":22103,"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 wavelet-domain hybrid model samples high-quality images in 16 steps","keywords":["wavelet-domain diffusion","score-based generative models","multi-scale adversarial learning","fast sampling","image generation","condition number","non-Gaussian wavelet coefficients","generative adversarial networks"],"falsifier":"Measure the empirical condition number and KL divergence to Gaussian of the coarsest LL band for a candidate dataset, then train WMGM on a version of that dataset whose LL band has been deliberately made non-Gaussian (for example by applying a heavy-tailed intensity transform before wavelet decomposition). If the FID gap over SGM and WSGM at 16 steps collapses, or the required step count jumps while the baselines move less, the Gaussian-tendency assumption is the load-bearing premise.","tokens_in":18758,"feed_emoji":"🖼️","tokens_out":11001,"duration_ms":95939,"temperature":0.7,"pith_summary":"The paper sets out to show that the slowness of diffusion-based image generation comes from forcing one Gaussian noising process onto the whole image, and that splitting the image by frequency fixes it. Its proposed model, the Wavelet Multi-scale Generative Model (WMGM), runs score-based diffusion only on the coarsest low-frequency wavelet band, where coefficients are close to Gaussian and well-conditioned, and uses a multi-scale adversarial network shared across scales to synthesize the sparse high-frequency bands from that low-frequency condition. On CelebA-HQ, AFHQ-Cat, and Colon at 128 × 128, the paper reports that WMGM with 16 total sampling steps reaches lower FID than both the spatial score baseline SGM and the wavelet score baseline WSGM, while using 89M parameters and 2.31 seconds per 100 images, compared with 160M/351M parameters and 15.09/11.10 seconds for those baselines. If the claim holds, fast sampling no longer has to trade away image quality, because the expensive iterative denoising is confined to a small well-conditioned subband.","feed_headline":"16-step wavelet hybrid beats score baselines at lower cost","feed_subtitle":"Diffusion runs on the low-frequency band alone; a shared GAN draws the sparse details.","key_machinery":"The load-bearing object is the orthogonal discrete wavelet transform as a coordinate change that splits image information by frequency. Because the wavelet matrix $A$ satisfies $AA^T = I$, the forward noising process and its reverse are dual between spatial and wavelet domains, so the score-based SDE can be run on wavelet coefficients without reformulating the dynamics. The mechanism then discriminates by band: the deepest low-frequency band is whitened by repeated averaging, making its covariance nearly isotropic with a small condition number, so the discretization bound $N \\le \\varepsilon^{-2} \\kappa^3$ implies a short reverse process suffices; the high-frequency bands are sparse and long-tailed, so instead of diffusion the paper applies Multi-Scale Adversarial Learning (MSAL), a 5-level U-Net generator with attention gates that maps each LL band plus injected noise to the corresponding high-frequency bands, trained with a Wasserstein adversarial loss combined with MSE and SSIM terms and with parameters shared across scales.","core_discovery":"The central claim is that wavelet-domain score-based diffusion is not the whole answer: the high-frequency bands are too sparse and non-Gaussian for Gaussian diffusion, while the low-frequency band is where diffusion works best. WMGM therefore factorizes the target distribution as $p(x_0) = \\prod_{k=1}^{S} p(x_H^k \\mid x_L^k) \\, p(x_L^S)$, generating the coarsest LL band $x_L^S$ by a short score-based reverse process and each set of high-frequency bands $x_H^k$ by a single adversarial pass of a shared multi-scale generator conditioned on $x_L^k$. The paper argues theoretically that the orthogonal wavelet transform maps the Ornstein-Uhlenbeck noising process and its reverse SDE onto the wavelet domain unchanged (duality), that iterated averaging whitens the LL band so its covariance has a low condition number and needs fewer discretization steps, and that the conditional high-frequency distribution remains sparse at deep scales, justifying a non-Gaussian model. Experimentally, the paper's Table 1 shows WMGM at 16 total sampling steps achieves FID 30.58/25.38 on CelebA-HQ (5K/30K), 16.29 on AFHQ-Cat, and 45.76 on Colon, each better than SGM and WSGM, with 89M parameters and 2.31 seconds per 100 images; the advantage persists from 6 to 192 steps in Figure 4.","pith_inferences":["Editorial inference: the same low-frequency diffusion plus sparse-detail adversarial split should transfer to other orthogonal or near-orthogonal multiscale decompositions such as Laplacian pyramids or block DCTs, since the paper's argument relies on band sparsity and whitening more than on wavelet-specific structure.","Editorial inference: a direct stress test is to measure the deepest LL-band condition number and KL divergence to Gaussian for each new dataset and check whether the optimal number of diffusion steps tracks those quantities; the paper's Appendix E.1 computes such diagnostics for two datasets but does not connect them to observed step counts.","Editorial inference: because the generator maps LL plus noise to high frequencies in one pass, the model may underrepresent multimodal detail textures even when FID improves; a perceptual-diversity metric computed across the injected noise for a fixed LL band would test that component.","Editorial inference: the reported speed advantage is measured per 100 images at a fixed resolution on one GPU type; at higher resolutions the cost shifts toward the adversarial generator, so the relative gain over full-image diffusion may grow or shrink depending on which component dominates."],"forward_implications":["At 16 total sampling steps, WMGM reports lower FID than SGM and WSGM on CelebA-HQ, AFHQ-Cat, and Colon, so the hybrid is the best of the three under the same step budget.","Because diffusion runs only on the coarsest LL band, the number of sampling steps needed for good quality is no longer tied to the full-image score condition number, which is the paper's explanation for the speedup.","Sharing generator parameters across wavelet scales reduces the model to 89M parameters, smaller than either baseline, so the added GAN component does not increase model size.","The ranking holds across total step counts from 6 to 192 and across face, animal, and histopathology datasets, indicating the frequency-split design transfers beyond a single domain.","The multi-scale adversarial variant outperforms single-scale, operator-learning, and standard-regression variants for the low-to-high-frequency mapping with half the parameters, so parameter sharing across scales is itself part of the reported quality gain."],"supporting_citations":[{"why":"Supplies the wavelet-domain score-based SGM baseline (WSGM) and the ill-conditioned-score analysis this paper extends.","marker":"[19]"},{"why":"Provides the non-Gaussian discretization bound (Theorem 6) used to justify wavelet-domain score modeling for natural images.","marker":"[18]"},{"why":"Defines the score-based SDE framework whose reverse process generates the coarsest low-frequency band.","marker":"[43]"},{"why":"Underlies the denoising score-matching loss and reverse discretization used for the low-frequency diffusion.","marker":"[21]"},{"why":"Introduces adversarial training, the basis for the multi-scale adversarial learning component.","marker":"[15]"},{"why":"Supplies the Wasserstein distance objective used in the MSAL generator and discriminator losses.","marker":"[1]"},{"why":"Establishes the power-law decay of natural-image spectra, the premise behind ill-conditioned spatial scores.","marker":"[12]"},{"why":"Defines the FID metric used to compare generation quality in the main experiments.","marker":"[20]"}],"fun_headline_variants":["Diffusion on low, GAN on high: wavelet hybrid wins at 16 steps","Wavelet split: score on smooth, GAN on detail, faster and cheaper","Multi-scale wavelet model beats diffusion baselines with fewer steps","WMGM: short diffusion for LL, GAN for detail, 16-step FID 30.58","Hybrid wavelet generative model: diffusion for coarse, GAN for fine"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method depends on the lowest-frequency wavelet band being close to Gaussian and well-conditioned enough that a short score-based diffusion can capture the global image structure; if that band stays heavy-tailed or ill-conditioned on a new dataset, the diffusion stage needs many more steps and the speed advantage disappears.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion on low, GAN on high: wavelet hybrid wins at 16 steps","Wavelet split: score on smooth, GAN on detail, faster and cheaper","Multi-scale wavelet model beats diffusion baselines with fewer steps","WMGM: short diffusion for LL, GAN for detail, 16-step FID 30.58","Hybrid wavelet generative model: diffusion for coarse, GAN for fine"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1491,"prompt_tokens":978,"completion_tokens":513,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":407}},"tokens_in":594,"tokens_out":513,"duration_ms":5203,"temperature":1.0,"reasoning_tokens":407,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:43:20.529718+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the empirical condition number and KL divergence to Gaussian of the coarsest LL band for a candidate dataset, then train WMGM on a version of that dataset whose LL band has been deliberately made non-Gaussian (for example by applying a heavy-tailed intensity transform before wavelet decomposition). If the FID gap over SGM and WSGM at 16 steps collapses, or the required step count jumps while the baselines move less, the Gaussian-tendency assumption is the load-bearing premise.","supporting_citations":[{"cited_title":"Wavelet score-based generative modeling","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-domain score-based SGM baseline (WSGM) and the ill-conditioned-score analysis this paper extends."},{"cited_title":"Wavelet score-based generative modeling","cited_arxiv_id":null,"evidence_quote":"Provides the non-Gaussian discretization bound (Theorem 6) used to justify wavelet-domain score modeling for natural images."},{"cited_title":"Wasserstein gan, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the Wasserstein distance objective used in the MSAL generator and discriminator losses."},{"cited_title":"Relations between the statistics of natural images and the response properties of cortical cells","cited_arxiv_id":null,"evidence_quote":"Establishes the power-law decay of natural-image spectra, the premise behind ill-conditioned spatial scores."}],"review_version":1}