{"id":"0edb4666-6133-4b7f-96e6-0bc2de8a39b4","arxiv_id":"2505.07866","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of DDPM, LDM, and WDM diffusion models for medical imaging, organized around training and inference efficiency.","lead":"A review paper compares three AI image generation approaches for medical scans, focusing on speed and computational cost. It claims to fill a gap in existing surveys by centering on large, high-resolution medical images.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim that it is the first to analyze latency for WSI/3D MRI/CT is contradicted by its own cited survey [11] and not delivered in its tables.","rationale":"I focus on the paper's stated reason for existing, rather than on individual table defects. The reader's weakest assumption (Table 3 incomparability) is legitimate and would already undermine the Section 4 recommendations, but the paper fails one step earlier: the claimed gap is contradicted by its own reference list, and the body does not perform the latency analysis for WSI/3D MRI/CT that the gap statement promises. The paper does assemble a large bibliography and organizes many applications of DDPM, LDM, and WDM; that organizational work has some standalone value. However, the central claim is explicitly a novelty claim about latency/time complexity for extremely large medical data, and that claim is not established. The concrete check above settles it by checking the cited prior survey and scanning for the promised measurements. Since this is the load-bearing premise, the reader's REJECT verdict should stand, although my stated weakest spot differs from the reader's Table 3 concern.","tokens_in":37336,"tokens_out":5577,"duration_ms":55225,"concrete_test":"Read Ulhaq et al. [11] and record whether it discusses inference latency, sampling speed, or computational cost of diffusion models; then scan Table 3 and Sections 3.1.2-3.3 of this manuscript for any latency/time value measured on WSI or 3D MRI/CT. If [11] covers latency and no such value appears, the uniqueness claim and the promised gap-filling analysis are both false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the Section 1 gap statement: 'no existing studies specifically discuss the latency and time complexity of generative models, specifically dealing with extremely large high-resolution data such as WSI or 3D MRI/CT,' with this survey filling that gap. This requires two things: the gap must exist, and the survey must actually deliver latency/time-complexity analysis for those data. Both conditions fail. First, the paper cites [11], 'Efficient Diffusion Models for Vision: A Survey,' and marks it in Table 1 as covering efficient sampling and fast convergence; a prior survey explicitly on efficient diffusion models makes the uniqueness sentence false as written. Second, the only quantitative efficiency table, Table 3, compares natural-image benchmarks (CIFAR-10, FFHQ, CelebA-HQ, LSUN-Church), not WSI or 3D MRI/CT, and the medical application sections (3.1.2, 3.2, 3.3) report no latency or time-complexity measurements for large high-resolution medical data. The promised gap-filling content is absent, so the central contribution is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a survey of diffusion-based generative models in medical imaging, organized around three model families: Denoising Diffusion Probabilistic Models (DDPM), Latent Diffusion Models (LDM), and Wavelet Diffusion Models (WDM). It claims to be the first survey to analyze latency and time complexity for extremely large high-resolution medical data such as whole slide images and 3D MRI/CT, and it promises a comparative analysis of generative quality and computational cost. The paper provides an overview of the three frameworks, applies them to natural and medical imaging tasks (generation, translation, reconstruction, segmentation, classification, editing, denoising, registration), and concludes with model-selection guidance: DDPM for maximum quality at low speed, LDM for a quality/speed balance, and WDM for fast inference with detail preservation.","tokens_in":37495,"tokens_out":5693,"duration_ms":52970,"significance":"The topic is timely, and the organizational effort is a genuine service to the community: the paper compiles a large number of recent applications into a structured taxonomy (Figure 5; Tables 4-9) and identifies three efficiency-relevant model families. This bibliography alone could be useful to a newcomer. However, the paper ships no reproducible experiments, no machine-checked proofs, and the quantitative comparisons are not derived from a unified protocol. More importantly, the central positioning claim—that no prior survey discusses latency for WSI/3D MRI/CT and that this survey fills that gap—is contradicted by the manuscript's own cited prior survey [11], and the promised latency analysis for large medical data is not actually delivered. As a review, the paper's value depends on the reliability of its comparative statements; the issues identified below compromise that reliability in load-bearing places.","major_comments":[{"comment":"The central gap statement—'no existing studies specifically discuss the latency and time complexity of generative models, specifically dealing with extremely large high-resolution data such as WSI or 3D MRI/CT'—is contradicted by the paper's own Table 1, which cites [11] 'Efficient Diffusion Models for Vision: A Survey' and marks that prior work as covering efficiency, efficient sampling, and fast convergence. Moreover, the promised gap-filling content is not present in the manuscript: Table 3 evaluates natural-image benchmarks only (CIFAR-10, FFHQ, CelebA-HQ, LSUN-Church), and Sections 3.1.2, 3.2, and 3.3 report no latency or time-complexity measurements for WSI or 3D MRI/CT. The claimed contribution is therefore unsupported by the manuscript's own evidence.","section":"Section 1 and Table 3"},{"comment":"The model-selection conclusions in Section 4 ('LDM balances quality and efficiency', 'WDM is a strong option if faster inference is required') rest on Table 3, where FID and inference-time values are collected from separate papers on different datasets with no error bars, no hardware specification, no sampling-step count, and no measurement protocol. For example, WDM inference times are 0.08 s on CIFAR-10 and 1.54 s on LSUN-Church, while LDM times are 0.43 s on FFHQ and 6.80 s on LSUN-Church; these values are not comparable across models because dataset, resolution, and hardware differ. This table cannot support the comparative recommendations advanced in Section 4.","section":"Table 3 and Section 4"},{"comment":"Equation (5) is mathematically garbled: it mixes '-log p_theta(x0|x1)', an undefined term 'KL(p_theta(x_S|x0) || pi(x_S) + (x,z))', and 'log p_theta(x)', so the variational lower bound is not correctly stated. Equations (1) and (4) further use inconsistent notation, with K used for a normal distribution, beta_t as variance in (1) but beta^2 I in (4), and the overline-alpha notation defined inconsistently with the sum/product indices. Because Section 2 is the foundation for the efficiency discussion that follows, this technical exposition needs a thorough rewrite.","section":"Section 2.1.1, Eq. (5)"},{"comment":"Section 2.2.6 states 'Generally, LDMs are considered more efficient for large image generation', whereas Table 2 assigns 'Generating large image' as High for WDM, Moderate for LDM, and Low for DDPM. This internal contradiction is not reconciled and is directly relevant to the model-selection guidance in Section 4, further undermining the comparative analysis.","section":"Section 2.2.6 vs Table 2"},{"comment":"The Conclusion states 'the foundational aspect of DDPMs [36]', but reference [36] is Lugmayr et al.'s RePaint inpainting paper, not Ho et al.'s DDPM paper, which is reference [15]. In a survey whose currency is the accurate mapping of literature, this is a load-bearing citation error that reduces the reader's confidence in the reference list as a whole.","section":"Section 6, citation [36]"}],"minor_comments":[{"comment":"'RePain [36] addressed form inpainting [35, 80]' contains two typos: the method name is RePaint, and 'form' should be 'from'.","section":"Section 3.1.1.3"},{"comment":"'LSUN-Chruch' is misspelled; it should be 'LSUN-Church'.","section":"Table 3"},{"comment":"The heading 'Challanges' is a typo; it should be 'Challenges'.","section":"Section 5.1 heading"},{"comment":"The roman-numeral reference mapping is ambiguous: '(iii)' is used both for Wavelet Diffusion and in the Datasets line, and '(xxiv)' is used for both Medical Diffusion and Wavelet SGM, each with different references; please assign unique numerals per reference.","section":"Figure 5"},{"comment":"The expectation notation E_{epsilon(x)} is nonstandard; the expectation should be over the data x, noise epsilon, and timestep t. Equation (8) also inconsistently writes tau_theta(x) in the condition and tau_theta(x') in the explanation.","section":"Eqs. (6)-(8)"},{"comment":"The note under Table 1 says 'The indicates that all aspects are covered, the shows...', but the symbols are not rendered; the legend is unreadable and should be fixed.","section":"Table 1 note"}],"recommendation":"reject","confidential_remarks":"The false-uniqueness claim is the paper's framing device and is contradicted by a survey the authors themselves cite and mark as covering efficient sampling. The quantitative centerpiece (Table 3) is not a valid comparison, and the foundational equations contain substantive errors rather than mere typos. A major revision would need to reframe the paper's contribution, replace or heavily caveat Table 3, and rewrite the technical sections; in my assessment this exceeds what minor or even major revision typically requires, so recommendation is reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about the diffusion-model survey (arXiv:2505.07866). My take: the useful core is a compact categorization of DDPM, LDM, and WDM with a decent map of recent applications, especially in medical imaging. The application tables (Tables 4-9) and Fig. 5 give a quick entry point to the literature. If someone wants names and citations for wavelet-based diffusion in medical imaging, this is a reasonable starting place.\n\nThe soft spots are not minor. The central uniqueness claim in Section 1—\"no existing studies specifically discuss the latency and time complexity of generative models, specifically dealing with extremely large high-resolution data such as WSI or 3D MRI/CT\"—is contradicted by reference [11], an efficient-diffusion survey they themselves cite and mark in Table 1 as covering efficient sampling and fast convergence. That alone would be a problem, but the survey also never delivers the promised latency analysis for WSI or 3D MRI/CT. Table 3 compares FID and inference times across different datasets (CIFAR-10, FFHQ, CelebA-HQ, LSUN-Church) with no hardware, measurement protocol, or error bars, so the Section 4 model-selection recommendations rest on numbers that are not comparable. Eq. (5) is garbled—it does not resemble a recognizable ELBO/kl term—and the conclusion cites [36] as the DDPM foundation when that reference is RePaint; the actual DDPM paper is [15]. These are not cosmetic typos; they affect whether the survey can be trusted as a map.\n\nThat said, the paper is not useless. As an orientation for someone new to the area, the three-way taxonomy and the medical applications survey have some value. But a review's value is its accuracy, and the inaccuracies here are load-bearing: the gap claim is false, the quantitative comparison is apples-to-oranges, and the equations are corrupted. I would not send this to a serious referee in its current form. It needs a major overhaul: fix the equations, correct the citations, rework or substantiate the uniqueness claim, and replace Table 3 with a controlled comparison or a clearly labeled reporting of published numbers with dataset and hardware context.\n\nMy recommendation: desk reject, but invite resubmission if the authors address these issues. Not worth reading-group time until then.","headline":"A survey with a useful categorization that is undermined by a false uniqueness claim, incomparable benchmark data, and corrupted equations; not ready for peer review.","tokens_in":38011,"tokens_out":1620,"would_cite":false,"duration_ms":17454,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This survey claims that no prior review has compared diffusion models on latency for gigapixel or 3D medical images, and positions WDM and LDM as the efficient answers.","keywords":["diffusion models","medical imaging","latent diffusion model","wavelet diffusion model","inference time","computational efficiency","generative models","survey"],"falsifier":"Run DDPM, LDM, and WDM on the same machine and the same dataset (for example, CIFAR-10 or CelebA-HQ), measure FID and wall-clock sampling time with identical hardware and hyperparameters, and check whether WDM remains the fastest and LDM the balanced middle ground from the table.","tokens_in":37132,"feed_emoji":"🩻","tokens_out":6402,"duration_ms":58739,"temperature":0.7,"pith_summary":"This review claims that no earlier survey has directly compared diffusion models on latency and time complexity for extremely large medical images such as whole-slide histopathology and 3D MRI/CT. To fill that gap, it organizes the field into three model families — DDPM, LDM, and WDM — and compares them on image quality, stability, convergence, speed, and scalability. The survey's practical conclusion is that DDPM gives the highest quality but slowest sampling, LDM balances quality and speed by working in a compressed latent space, and WDM is fastest by diffusing in the wavelet domain. It matters because medical imaging needs fast, reliable generation at very high resolutions, where inference time is a real bottleneck. The paper argues that model selection should be application-specific rather than a single universal choice.","feed_headline":"New survey ranks diffusion models by speed on huge medical images","feed_subtitle":"Weighs DDPM's quality, LDM's balance, WDM's speed for whole-slide and 3D imaging.","key_machinery":"The organizing machinery is a three-way taxonomy — DDPM, LDM, WDM — mapped against the generative learning trilemma (sample quality, sampling speed, and mode coverage/diversity), with comparisons carried by quality metrics such as Fréchet Inception Distance (FID), parameter counts, and wall-clock inference times. Within the taxonomy, the load-bearing mechanisms are the LDM's pre-trained autoencoder that shifts diffusion from pixel space to latent space, and the WDM's discrete wavelet transform that decomposes an image into four quarter-resolution subbands, reducing the diffusion computation by a factor of four. These two mechanisms are what the survey credits with closing the efficiency gap left by DDPM's step-by-step denoising.","core_discovery":"The paper's central claim is that a gap exists in the diffusion-model literature: existing surveys cover quality, applications, and general computational cost, but none specifically discusses latency and time complexity for extremely large high-resolution data such as whole-slide images or 3D MRI/CT. On the paper's own terms, filling that gap yields a consistent quality–speed–scalability ranking: DDPM leads in fidelity at the cost of slow iterative sampling (e.g., 80.5 seconds on CIFAR-10), LDM operates in the latent space of a pre-trained autoencoder to reach a middle ground, and WDM decomposes images with a discrete wavelet transform, cutting dimensionality fourfold and delivering the fastest reported times. This ordering supports the review's claim that WDM and LDM address the diffusion trilemma, and that applications requiring fine detail at speed — whole-slide pathology, 3D CT/MRI — are exactly where the choice of model matters most.","pith_inferences":["Beyond the paper: a standardized benchmark that runs DDPM, LDM, and WDM on the same hardware, same dataset, and same resolution (ideally whole-slide images or 3D MRI/CT) would turn the survey's cross-paper numbers into a testable ranking.","Beyond the paper: because LDM loses fine details and WDM is moderate in sampling efficiency, a hybrid that first compresses into a latent space and then diffuses in the wavelet domain is a natural candidate for pushing both speed and texture fidelity.","Beyond the paper: for real-time clinical deployment, the survey's own ranking implies that DDPM would be used offline while WDM or distilled one-step variants are the route to interactive latency, an extension the review does not explore.","Beyond the paper: the reported gap also suggests that future medical-imaging surveys should report hardware, measurement protocol, and error bars so latency claims become reproducible rather than aggregated."],"forward_implications":["For tasks where generation quality dominates and time is not critical, DDPM remains the strongest choice; on CIFAR-10 it reaches the best FID (3.21) but takes about 80.5 seconds.","For applications that need both quality and speed, LDM is the recommended middle ground, achieving competitive FID (4.98 on FFHQ) at about 0.43 seconds.","For speed-critical, detail-heavy cases such as whole-slide pathology, WDM is positioned as the fastest option (0.08 seconds on CIFAR-10 with FID 4.01) and best at preserving texture.","The survey claims to fill a gap no existing survey covers: latency and time complexity for extremely large high-resolution data such as whole-slide images and 3D MRI/CT.","Model selection should be application-specific rather than a single universal best model."],"supporting_citations":[{"why":"Introduces DDPM, whose slow iterative denoising is the efficiency baseline this survey compares against.","marker":"[15]"},{"why":"Introduces the latent diffusion model, the middle-ground efficient alternative the survey recommends for balanced quality and speed.","marker":"[14]"},{"why":"Introduces wavelet diffusion, the fastest family in the comparison, central to the survey's latency discussion.","marker":"[58]"},{"why":"Frames the generative trilemma (quality, sampling speed, convergence) that organizes the survey's evaluation criteria.","marker":"[65]"},{"why":"Earlier medical-image diffusion survey that, the paper argues, omits latency and time complexity, establishing the gap being filled.","marker":"[12]"},{"why":"Shows gigapixel-scale whole-slide histopathology generation, the extreme-resolution case where latency matters most.","marker":"[51]"},{"why":"Wavelet score-based model that keeps sampling iterations independent of image size, supporting WDM's scalability claim.","marker":"[70]"}],"fun_headline_variants":["Speed rank of diffusion models for huge medical images","Which diffusion model is fastest for medical scans?","DDPM vs LDM vs WDM: speed for large medical images","Review: fastest diffusion model for whole-slide and 3D imaging","Efficient diffusion models for medical imaging: speed ranking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recommendation that LDM is balanced and WDM is fast depends on comparing FID and inference times reported in separate papers, on different datasets, with no shared hardware or error bars; if those numbers are not directly comparable, the ranking loses its support.","fun_headline_variants_meta":{"raw":{"variants":["Speed rank of diffusion models for huge medical images","Which diffusion model is fastest for medical scans?","DDPM vs LDM vs WDM: speed for large medical images","Review: fastest diffusion model for whole-slide and 3D imaging","Efficient diffusion models for medical imaging: speed ranking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1405,"prompt_tokens":938,"completion_tokens":467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":554,"tokens_out":467,"duration_ms":4331,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:53:23.054538+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run DDPM, LDM, and WDM on the same machine and the same dataset (for example, CIFAR-10 or CelebA-HQ), measure FID and wall-clock sampling time with identical hardware and hyperparameters, and check whether WDM remains the fastest and LDM the balanced middle ground from the table.","supporting_citations":[],"review_version":1}