{"id":"88e99ab6-f167-438f-ad45-8543af261477","arxiv_id":"2411.12181","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A beta noise schedule, a sinusoidal curriculum, and weighted attention gates improve one-step consistency model training for images and low-dose CT denoising.","lead":"This paper adds two tweaks to consistency models, a fast type of AI image generator, and tests them on low-dose CT scans: a beta-shaped noise schedule and a sine-wave training curriculum. The authors report better one-step image quality on CIFAR-10 and CelebA, and improved denoising of low-dose CT slices.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The LDCT headline result rests entirely on Table 3, which is labeled 'copied from other paper' with no citation; unless those baseline numbers were produced under exactly the same evaluation protocol, the claimed PSNR/LPIPS advantage is not established.","rationale":"The reader identified hyperparameter tuning as the weakest assumption, and that is a real in-sample selection issue for the unconditional results. However, the single most load-bearing concern for the paper's stated purpose is the provenance of the LDCT comparison in Table 3. The table caption itself says 'copied from other paper' without citing the source, so the baseline numbers and evaluation protocol are unverifiable. The conditional model is image-conditioned, while several baselines are unconditional or at different NFEs, so a mismatch in data split or preprocessing could easily produce the observed differences. The paper also provides no code, data, or formal verification, so this table is the only evidence for the central LDCT claim. Reinforcing the reader's CONDITIONAL verdict is appropriate: the paper should be accepted only if the authors provide a cited, reproducible comparison and a shared evaluation script.","tokens_in":9856,"tokens_out":6936,"duration_ms":71075,"concrete_test":"Request the source of Table 3 and the authors' exact validation protocol; then run the official baseline implementations (DDPM, EDM, DDIM, CD, PFGM++, PS-PFCM) on the same Mayo validation slices with the same preprocessing and metric script used for HN-iCT-CN. If any reproduced baseline PSNR or LPIPS differs from the table by more than 1 dB or 0.01, the claimed advantage is an artifact of the copied comparison rather than a real result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's title and main contribution is low-dose CT enhancement, and the sole quantitative evidence is §4.3 Table 3. The caption explicitly states 'Quantitative Results, copied from other paper' with no reference given. Consequently, the listed baseline numbers (DDPM, EDM, DDIM, CD, PFGM++, PS-PFCM) cannot be checked against the proposed HN-iCT-CN: their data split, CT preprocessing (window/level, clipping), slice selection, evaluation mask, and metric implementation may all differ. The proposed model is an image-conditioned single-step consistency model, whereas several baselines use different NFEs and possibly different conditioning; without a controlled protocol, the 'improvement' could reflect evaluation mismatch rather than algorithmic gain. The reported drop in LPIPS from 0.152 to 0.016 and PSNR from 41.50 to 44.07 is unusually large for a single-step method and requires independent reproduction. This missing provenance is the load-bearing weakness: if the table is not a faithful comparison, the central claim of the paper is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two modifications to consistency-model training: a beta noise distribution for sampling noise levels and a sinusoidal curriculum for increasing the number of noise levels during training. Together these form 'High Noise Improved Consistency Training' (HN-iCT), and when combined with a conditional architecture using 'Weighted Attention Gates' (WAG), the method is applied to low-dose CT denoising (HN-iCT-CN). The authors report that with NFE=1, HN-iCT achieves FID 10.50 on CIFAR10 and 12.31 on CelebA using a smaller model than the CT/iCT baselines, and that HN-iCT-CN achieves PSNR 43.68-44.07 and LPIPS 0.016 on the Mayo LDCT validation set, outperforming several diffusion baselines. The paper also includes an ablation study separating the effects of curriculum type and noise distribution.","tokens_in":10136,"tokens_out":4530,"duration_ms":47818,"significance":"If the reported results are reproducible and the comparisons are fair, the paper would make a useful contribution: single-step consistency models with improved noise scheduling are of interest to the generative-modeling community, and fast LDCT denoising has clear clinical motivation. The authors are to be credited for evaluating against publicly available CT/iCT checkpoints, for providing an ablation study, and for examining the effect of high-noise-level weighting. However, the significance is conditional on resolving two load-bearing issues: the provenance and protocol of the LDCT comparison table, and the separation between hyperparameter tuning and reported evaluation. The CIFAR10 and CelebA results are also weakened by the apparent use of the same dataset for tuning and headline reporting.","major_comments":[{"comment":"Table 3 is captioned 'Quantitative Results, copied from other paper' and provides no citation or specification of which entries were copied. Since the central LDCT claim of the paper rests entirely on this table, the listed baseline numbers (DDPM, EDM, DDIM, CD, PFGM++, PS-PFCM) cannot be verified to use the same data split, CT window/level preprocessing, slice selection, evaluation mask, or metric implementation as the HN-iCT-CN entries. The unusually large LPIPS reduction from 0.152 to 0.016 and the absence of a controlled protocol make this a load-bearing unsupported comparison. The authors must either reproduce the baselines under the identical evaluation protocol or provide a precise citation and explicitly state which rows were copied and which were newly computed, together with complete evaluation details.","section":"§4.3, Table 3"},{"comment":"The hyperparameters appear to be tuned on the same datasets used to report the headline numbers. Section 3.2 contains the note 'CLAIM and modify: The set of parameters were empirically determined by using CIFAR10', and Section 4.2 reports CIFAR10 FID with α=1.5, β=5. Similarly, Section 4.3 states that α=0.5 was selected based on experiments on the LDCT validation set, and Table 3 reports results on that same validation set. Unless the authors document a clear separation between tuning and evaluation—for example, a validation split used for selection and a held-out test split for final numbers—the reported FID and PSNR improvements are in-sample fits rather than predictive comparisons.","section":"§3.2, §4.2, §4.3"},{"comment":"The proposed sinusoidal curriculum as written in Eq. (9) is non-monotonic: for k in [0,K], the term |sin(π·3·k/(2K))| reaches 1 at k=K/3, returns to 0 at k=2K/3, and then rises again to 1 at k=K. Thus N(k) increases, then decreases back to s0, then increases again. This is inconsistent with the text's description of a curriculum that 'gradually decreases the rate at which the number of noise varieties increases across levels.' The authors should clarify whether the intended schedule is monotonic and provide the exact discrete implementation used in training.","section":"§3.2, Eq. (9)"},{"comment":"Table 4's caption states 'β=1.5, α=5 for Beta distribution', whereas Table 2 and Section 4.3 use α=1.5/β=5 and α=0.5/β=5, respectively. Because the beta shape parameters directly control the noise-level weighting, this inconsistency prevents a reader from comparing the ablation results with the main reported results. The values must be corrected and the notation made consistent throughout.","section":"§4.4, Table 4"},{"comment":"The paper attributes the medical-image denoising improvement to the Weighted Attention Gate (WAG) modules, but no ablation isolates the contribution of WAG. All HN-iCT-CN rows in Table 3 use WAG, so the reported PSNR/LPIPS gains cannot be attributed to the WAG mechanism rather than to the beta noise distribution, the sinusoidal curriculum, or the conditional architecture as a whole. An ablation with and without WAG under the same training budget is needed to support the conclusion.","section":"§3.1, §4.3, §5"}],"minor_comments":[{"comment":"Equation numbers are duplicated: Eq. (5) is used for both the consistency matching loss and the log-normal noise distribution, and Eq. (7) is used for both the EMA update and the pseudo-Huber loss. Please renumber.","section":"§2, equations"},{"comment":"The text says 'all models compared in table 2 are trained locally' but also refers to 'publicly available sources' and 'versions from v1 to v3.' Please clarify which rows are locally trained, which are taken from public checkpoints, and what training hyperparameters were used for the locally trained baselines.","section":"§4.2"},{"comment":"References [6] and [7] are the same Karras et al. paper and should be merged or distinguished.","section":"References"},{"comment":"The sentence 'comprehends iAG modules' appears to be a typo; it should likely read 'comprises' or 'includes.'","section":"§3.1"},{"comment":"The discussion of the α parameter states that 'a reduction in batch size leads to a decrease in α,' but no quantitative evidence or experimental protocol is given. Please either provide the supporting experiments or state this as a hypothesis.","section":"§4.3"},{"comment":"Table 5 lists FID values for models trained with high-noise-level ratios, but the text says 'adding high level noise at 10% of the mini-batch length has effects on denoising performance conversely' without reporting a 10% row. Clarify whether the 10% experiment is omitted or reported elsewhere.","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is not yet in a publishable state: the LDCT evidence base has a missing provenance, the hyperparameter tuning appears to overlap with the reported evaluation, and there are internal inconsistencies in the notation and equations. These issues are fixable in principle, so I recommend major revision rather than rejection. The editor may also wish to check that the 'CLAIM and modify' note in Section 3.2 is removed or properly rewritten, as its presence suggests the manuscript is incomplete."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the unconditional experiments are the only part that stands, and the medical claim in the title doesn't. Table 3, the sole evidence for LDCT denoising, is captioned \"copied from other paper\" with no citation. That means the PSNR/LPIPS numbers are not tied to a defined evaluation protocol; the baselines could be from a different split, pre-processing, or metric implementation. The 44 dB PSNR and 0.016 LPIPS are interesting, but they're not yet claims.\n\nWhat is new: beta noise scheduling, sinusoidal curriculum, and the WAG module. The ablation in Table 4 suggests both modifications help on CIFAR10 at 400K steps. The FID improvements in Table 2 are plausible: they train a 234M model and beat larger CT/iCT baselines at NFE=1. That's worth credit.\n\nWhere it's soft: alpha and beta were picked using CIFAR10, then Table 2 reports CIFAR10 FID; that's in-sample fitting, not a prediction. The conditional alpha=0.5 was chosen after validation-set experiments. The text even contains a \"CLAIM and modify\" note in Section 3.2, which says the parameters were empirically determined on CIFAR10. That's a red flag. There's also no code or data linked despite mentioning a public repo. The baseline models in Table 2 are trained locally, but the table doesn't give iteration counts, batch sizes, or training details for those runs. The conditional model uses 14M/22M parameters, but the comparison baselines in Table 3 are from different papers with unknown sizes and conditioning.\n\nThis paper is for researchers working on consistency training who want an alternative to log-normal noise; the beta schedule idea could be a building block. For the LDCT community it's not usable yet. It needs a full rewrite and a proper LDCT evaluation before it deserves referee time. I'd desk-reject in this form and invite a resubmission with a complete comparison table, held-out hyperparameter selection, and code release. For a reading group, it's a useful case study in evaluation hygiene.","headline":"The unconditional training tweaks are worth a look, but the LDCT table is an unsourced copy, so the paper's main claim is unverifiable as written.","tokens_in":10652,"tokens_out":3023,"would_cite":false,"duration_ms":31492,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that consistency models—single-step generative networks—improve markedly when trained with a beta-distributed noise schedule and a sinusoidal timestep curriculum, and that the same recipe, conditioned on low-dose CT…","keywords":["consistency models","single-step sampling","low-dose CT denoising","beta noise distribution","sinusoidal curriculum","weighted attention gates","noise scheduling","consistency training"],"falsifier":"A decisive check is to retrain with the reported recipe while keeping the evaluation sets untouched: pick $\\alpha$ and $\\beta$ on a CIFAR-10 validation split and measure test FID on data the tuning never saw, and likewise choose $\\alpha = 0.5$ and the WAG weight on a Mayo training subset before touching the validation slices; if the margins over iCT (13.50 to 10.50 on CIFAR-10) and over the listed CT baselines persist under that discipline, the claim is predictive. The table header 'copied from other paper' on the LDCT results means the baseline PSNR and LPIPS numbers should be regenerated under identical preprocessing and model configuration before the comparison is taken as settled. The ablation caption also writes $\\beta = 1.5$, $\\alpha = 5$, the reverse of the main text's $\\alpha = 1.5$, $\\beta = 5$, so the hyperparameters used in the ablation table need to be clarified.","tokens_in":9637,"feed_emoji":"🩻","tokens_out":9150,"duration_ms":78148,"temperature":0.7,"pith_summary":"Diffusion-based image generation is powerful but slow because it needs many network evaluations to denoise step by step; consistency models replace that loop with a single evaluation, and this paper claims their training procedure can be substantially improved. The authors introduce a $\\beta$ noise distribution, whose two shape parameters $\\alpha$ and $\\beta$ tilt the training batch toward high or low noise, combined with a sinusoidal curriculum that grows the number of timesteps smoothly, a recipe they call High Noise Improved Consistency Training (HN-iCT). At a single function evaluation, HN-iCT reports FID 10.50 on CIFAR-10 and 12.31 on CelebA with a 234M-parameter model, beating 897M-parameter CT and iCT baselines, and an image-conditioned variant with weighted attention gates reports PSNR 43.68–44.07 and LPIPS 0.016 on the Mayo low-dose CT validation set. If these numbers hold, clinical-grade low-dose CT enhancement would become a single network pass rather than a multi-step diffusion process, at a fraction of the compute.","feed_headline":"Consistency training with beta noise beats iCT in one step","feed_subtitle":"Beta noise scheduling and a sinusoidal curriculum improve single-step FID and denoise low-dose CT in one pass.","key_machinery":"Three mechanisms carry the argument. The $\\beta$ noise distribution, defined on $[0,1]$ with density $f(x;\\alpha,\\beta) = x^{\\alpha-1}(1-x)^{\\beta-1}/B(\\alpha,\\beta)$, replaces the log-normal schedule used by iCT; it lets the training batch be tilted toward high noise levels while keeping low noise heavily weighted. The sinusoidal curriculum sets the number of timesteps as $N(k) = \\min(\\lfloor s_1 \\sin(\\pi \\cdot 3k / 2K) + s_0 \\rfloor + 1, s_1 + 1)$ with $s_0 = 20$, $s_1 = 250$, growing the variety of noise levels smoothly so adjacent trajectory points remain close enough to learn consistent mappings. For the conditional model, the Weighted Attention Gate (WAG) squares the attention map and combines features extracted from the conditioned input with skip connections at a default weight of 0.8, so the low-dose CT image supervises the unconditional generation without dominating it. The machinery's job is to make the single-step consistency loss—the distance between the student network's outputs at adjacent noise levels—trainable along a well-populated, well-ordered trajectory.","core_discovery":"The central discovery claimed is that consistency training fails mainly because its noise distribution and curriculum misallocate training effort, and that both can be fixed without new network machinery. The authors report that mixing high noise levels into mini-batches—up to roughly 4% of samples drawn from noise levels 40 to 80—sharply improves denoising, and that the $\\beta$ distribution captures this effect with two tunable shape parameters ($\\alpha = 1.5$, $\\beta = 5$ for unconditional tasks; $\\alpha = 0.5$ for the conditional model). The sinusoidal curriculum replaces the improved curriculum's doubling of the timestep count every 50k steps with $N(k) = \\min(\\lfloor s_1 \\sin(\\pi \\cdot 3k / 2K) + s_0 \\rfloor + 1, s_1 + 1)$, with $s_0 = 20$, $s_1 = 250$, keeping trajectory points traceable as the schedule grows. Together these changes are claimed to be why a 234M-parameter HN-iCT out-scores 897M-parameter CT and iCT baselines at NFE=1 (FID 10.50 vs 14.32 and 13.50 on CIFAR-10), and why the conditioned HN-iCT-CN reduces LPIPS to 0.016 on low-dose CT. The ablation table makes the pairing direct: improved curriculum plus log-normal noise gives FID 21.19, improved plus $\\beta$ 17.31, sinusoidal plus log-normal 17.85, and the full sinusoidal plus $\\beta$ combination 10.51.","pith_inferences":["If the gain comes from covering high noise levels that the log-normal schedule underweights, a natural test is whether other heavy-tailed or mixture schedules (for example, log-normal plus a fixed mass at high $\\sigma$) match HN-iCT; the paper's own appendix experiments with manually mixing 4% high-noise samples suggest the mechanism, not the specific distribution, may be what matters.","The reported dependence of $\\alpha$ on batch size hints that the schedule is compensating for the effective number of high-noise samples per update; a follow-up could make $\\alpha$ a function of batch size rather than a tunable constant.","For clinical deployment, the meaningful comparison is against non-generative denoisers such as NLM, BM3D, and supervised CNNs as much as against diffusion baselines; the paper targets generative methods, and a dedicated clinical-reader study would be needed before drawing conclusions about diagnostic utility.","The WAG weighting (0.8) and the squaring of the attention map are described qualitatively; a systematic sweep of the weight parameter and attention sharpening could determine whether the reported PSNR gain is robust across patients and reconstruction kernels."],"forward_implications":["Single-step (NFE=1) consistency generation becomes competitive with much larger models: the same FID can be reached with roughly a quarter of the parameters of the CT and iCT baselines.","Low-dose CT enhancement can run in one function evaluation at 512×512 without resizing, with PSNR around 44 dB and LPIPS 0.016, above the listed multi-step diffusion baselines.","The $\\alpha$ parameter is tied to batch size: the paper reports that smaller batches require smaller $\\alpha$, giving a practical rule for transferring the noise schedule to new settings.","The two proposed components combine super-additively: sinusoidal curriculum plus beta noise (FID 10.51) beats either alone (17.85 and 17.31), so the claimed benefit is the pair, not either piece.","Because the model uses CT-style direct data learning with no pretrained teacher, the technique avoids distillation cost and is not capped by a teacher model's quality."],"supporting_citations":[{"why":"Supplies the improved consistency training (iCT) baseline—curriculum, log-normal noise schedule, and pseudo-Huber loss—that the paper modifies and compares against.","marker":"[15]"},{"why":"Provides the consistency model parameterization, the consistency training objective, and the single-step sampling framework that HN-iCT builds on.","marker":"[16]"},{"why":"Sets the noise discretization ($\\rho = 7$, $\\sigma_{\\min} = 0.002$, $\\sigma_{\\max} = 80$) used for the noise sequences and supplies the EDM baseline in the comparisons.","marker":"[6]"},{"why":"Source of the beta distribution as a flexible noise prior, which the paper adapts into its beta noise scheduling.","marker":"[22]"},{"why":"The Mayo Clinic low-dose CT grand challenge dataset used for training and validating the image-conditioned model.","marker":"[12]"},{"why":"The main diffusion baseline for low-dose CT denoising in the quantitative comparison table.","marker":"[19]"},{"why":"The attention gate mechanism from attention U-Net that the paper adapts into its Weighted Attention Gate modules.","marker":"[13]"},{"why":"The single-step consistency baseline (PS-PFCM) for low-dose CT denoising that HN-iCT-CN is compared with.","marker":"[4]"}],"fun_headline_variants":["Beta noise beats iCT in single-step CT enhancement","One-step CT: beta noise outperforms 4x larger models","High noise, sinusoidal curriculum fix one-step CT","Consistency training upgrade: beta noise for low-dose CT","Single-step denoising beats iCT with beta noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline gains rest on empirically tuned hyperparameters ($\\alpha = 1.5$, $\\beta = 5$, $s_0 = 20$, $s_1 = 250$, $\\alpha = 0.5$, WAG weight 0.8) that the paper states were determined by experiments on CIFAR-10, then applied without change to CelebA and the Mayo low-dose CT set; if those values were chosen with the evaluation sets in view, the reported FID and PSNR numbers measure in-sample fit rather than prediction.","fun_headline_variants_meta":{"raw":{"variants":["Beta noise beats iCT in single-step CT enhancement","One-step CT: beta noise outperforms 4x larger models","High noise, sinusoidal curriculum fix one-step CT","Consistency training upgrade: beta noise for low-dose CT","Single-step denoising beats iCT with beta noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000267,"raw_usage":{"total_tokens":1686,"prompt_tokens":1088,"completion_tokens":598,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":518}},"tokens_in":704,"tokens_out":598,"duration_ms":6426,"temperature":1.0,"reasoning_tokens":518,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T17:49:53.468999+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive check is to retrain with the reported recipe while keeping the evaluation sets untouched: pick $\\alpha$ and $\\beta$ on a CIFAR-10 validation split and measure test FID on data the tuning never saw, and likewise choose $\\alpha = 0.5$ and the WAG weight on a Mayo training subset before touching the validation slices; if the margins over iCT (13.50 to 10.50 on CIFAR-10) and over the listed CT baselines persist under that discipline, the claim is predictive. The table header 'copied from other paper' on the LDCT results means the baseline PSNR and LPIPS numbers should be regenerated under identical preprocessing and model configuration before the comparison is taken as settled. The ablation caption also writes $\\beta = 1.5$, $\\alpha = 5$, the reverse of the main text's $\\alpha = 1.5$, $\\beta = 5$, so the hyperparameters used in the ablation table need to be clarified.","supporting_citations":[{"cited_title":"Improved techniques for training consistency models","cited_arxiv_id":null,"evidence_quote":"Supplies the improved consistency training (iCT) baseline—curriculum, log-normal noise schedule, and pseudo-Huber loss—that the paper modifies and compares against."},{"cited_title":"Elucidating the design space of diffusion-based generative models","cited_arxiv_id":null,"evidence_quote":"Sets the noise discretization ($\\rho = 7$, $\\sigma_{\\min} = 0.002$, $\\sigma_{\\max} = 80$) used for the noise sequences and supplies the EDM baseline in the comparisons."},{"cited_title":"Beta diffusion, 2023","cited_arxiv_id":null,"evidence_quote":"Source of the beta distribution as a flexible noise prior, which the paper adapts into its beta noise scheduling."},{"cited_title":"McCollough, Adam C","cited_arxiv_id":null,"evidence_quote":"The Mayo Clinic low-dose CT grand challenge dataset used for training and validating the image-conditioned model."},{"cited_title":"Low-dose ct using denoising diffusion probabilistic model for 20× speedup, 2022","cited_arxiv_id":null,"evidence_quote":"The main diffusion baseline for low-dose CT denoising in the quantitative comparison table."},{"cited_title":"Attention u-net: Learning where to look for the pancreas, 2018","cited_arxiv_id":null,"evidence_quote":"The attention gate mechanism from attention U-Net that the paper adapts into its Weighted Attention Gate modules."},{"cited_title":"Poisson flow consistency models for low-dose ct image denoising, 2024","cited_arxiv_id":null,"evidence_quote":"The single-step consistency baseline (PS-PFCM) for low-dose CT denoising that HN-iCT-CN is compared with."}],"review_version":1}