{"id":"5a483dce-7e3e-4165-8a40-499ff294c1cd","arxiv_id":"2502.08598","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Constant total variance with an inverse-sigmoid SNR schedule improves few-step diffusion sampling, giving stable molecules in 4 steps.","lead":"The paper proposes a way to separate total noise variance and signal-to-noise ratio in diffusion models, and shows that keeping variance constant while tuning SNR improves few-step generation. It introduces a new SNR schedule, VP-ISSNR, that generates stable molecules in as few as 4 steps and matches a strong baseline on images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gains rest on sampling with models trained under other schedules, so the measured improvements could be artifacts of score-model bias rather than of the proposed TV/SNR schedule; a retrained VP-ISSNR model is the decisive missing check.","rationale":"The paper's central empirical claim is that constant-TV and VP-ISSNR schedules improve few-step sampling. The clean way to test a schedule is to train under that schedule; instead, the paper reuses a DDPM-cos-trained model (Appendix C) and pretrained EDM models, relying on the assertion that conditioning on scaled SNR makes the model schedule-agnostic. This is a plausible but unproven generalization assumption. The toy trajectory analysis in Figures 1B and A8, which uses exact scores and shows lower curvature for VP-ISSNR, is genuine supporting evidence for the geometry story, and Figure A4 shows that an EDM-trained model also performs well under VP-ISSNR sampling, weakening the strongest version of the training-artifact worry. However, neither piece of evidence replaces a direct VP-ISSNR retraining experiment. The magnitude of the molecular improvement makes this the pivotal uncertainty: if retraining changes the ranking, the central contribution is not what it appears. Missing error bars and per-NFE Bayesian tuning for images are secondary concerns; the fixed-parameter molecular results carry the main claim. The reader's weakest assumption points to exactly this issue, and the recommended conditional verdict remains appropriate. The work is internally consistent and the derivations check out, but the decisive empirical validation is outstanding.","tokens_in":24457,"tokens_out":10055,"duration_ms":118979,"concrete_test":"Train the same QM9 score architecture from scratch with the VP-ISSNR schedule (eta=1.0, kappa=2.0, tmin=0.01, tmax=0.99, tau(t)=1) using the Appendix C training protocol, then evaluate stability with Euler ODE at NFE in {4, 8, 16, 32, 64, 128} over the same 2.5k generated structures. Compare with Figure 3. If the VP-ISSNR-trained stability at 8 NFEs is more than 10 percentage points lower than the reported DDPM-cos-trained VP-ISSNR curve, the claimed schedule advantage is at least partly a training-schedule artifact; if it matches or exceeds the reported curve, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that a score network trained on one schedule can be reused to compare schedules. Appendix C states that the molecular model is always trained with DDPM-cos (nu=1) and tau(t)=1, conditioning only on a scaled SNR, and that this makes it compatible with various TV and SNR schedules during sampling without retraining. All VP-variant comparisons and the VP-ISSNR SOTA numbers in Figure 3 are generated with this single model, plus pretrained EDM models for images in Figure 4. If the network's score approximation is not uniformly accurate over the (x_hat, gamma) pairs queried by different schedules, schedules that emphasize gamma regions where the model is inaccurate will look worse for reasons unrelated to TV/SNR geometry. The paper's own Figure A4 partially addresses this by training an EDM-schedule model and showing that VP-ISSNR sampling still performs well, but it never trains with VP-ISSNR itself. The central claim that constant-TV and VP-ISSNR schedules improve sampling therefore remains entangled with the training schedule. A retraining experiment is the check that would settle whether the improvement is a property of the schedule or an artifact of how the reused model was trained.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a total-variance/signal-to-noise-ratio (TV/SNR) disentangled framework for diffusion noise schedules, in which the perturbation kernel is reparameterized by a TV-controlling function τ(t) and an SNR-controlling function γ(t). The authors derive forward/reverse SDEs for this parameterization, introduce constant-TV (VP) variants of existing VE schedules, and propose a new VP-ISSNR schedule based on the exponential of an inverse sigmoid SNR. They report improved few-step molecular structure generation on QM9 (stability rate as a function of NFE) and improved/competitive FID on CIFAR-10 and FFHQ, and supplement this with a numerical analysis of ODE trajectory curvature and marginal-support growth. The central claims are that exploding-TV schedules can be improved by constant-TV counterparts and that the proposed ISSNR schedule enables state-of-the-art few-step generation, especially for molecules.","tokens_in":24872,"tokens_out":12291,"duration_ms":122176,"significance":"If the claims hold, the TV/SNR reformulation is a useful unifying perspective, and the empirical gains in few-step molecular generation are practically significant. The manuscript ships an open-source implementation, recovers several existing schedules in a single table (Table A1), and includes a toy experiment on trajectory curvature and marginal support that goes beyond pure benchmarking. The VP-vs-VE comparisons on images and molecules use shared pretrained models, which is a reasonable first-order protocol, and Figure A4 provides a partial check of schedule transfer for molecules. However, the correctness of the derived SDE drift is compromised by an algebraic error, and the tuned image hyperparameters are selected per NFE without a stated validation protocol. These issues are load-bearing for the empirical conclusions, so the paper requires major revision before the central claims can be accepted.","major_comments":[{"comment":"The drift coefficient is derived incorrectly. For a^2 = τ^2γ^2/(1+γ^2), differentiating log a gives f(t) = d/dt log a = τ'(t)/τ(t) + γ'(t)/[γ(t)(1+γ^2(t))], not τ'(t)/τ(t) + γ'(t)/γ(t) * (1+γ^2(t)). The last algebraic step in the derivation inverts the factor (1+γ^2). Consequently, the forward SDE and the reverse ODE/SDE used in Sections 3 and 4 are not the SDEs that correspond to the perturbation kernel in Eq. (8). For the pure VE case τ=1, γ=1/σ, the correct drift is f=0, whereas the paper's formula gives f = -σ'/σ (1+1/σ^2), a spurious non-zero drift. Since all sampling experiments, including the VP-ISSNR results, use this ODE/SDE, the authors must correct the derivation and re-run the experiments with the correct drift, or demonstrate that the code implemented the correct drift despite the erroneous derivation.","section":"Appendix B.3, Eq. (42)"},{"comment":"The 'ISSNR[tuned]' results are obtained by Bayesian optimization of η and κ separately for each NFE (32 trials), but the manuscript does not specify the selection criterion or whether the FID used for selection is computed on the same test set used for the reported numbers. Tuning directly against the reported FID would make the low-NFE improvements optimistic and not directly comparable with fixed-schedule baselines such as EDM and OTFM, whose hyperparameters were set in prior work. Please state the selection protocol (e.g., a validation split), and present the fixed 'scaled' variant as the primary image claim, with the tuned variant clearly labeled as an upper-bound/exploratory result.","section":"Section 4.2 and Figure 4"},{"comment":"The headline molecular results are produced by a single model trained with the DDPM-cos (ν=1) schedule and then evaluated with other schedules, relying on the network being conditioned on SNR rather than time. Figure A4 shows that a model trained with the EDM SNR schedule also benefits from VP-ISSNR sampling relative to other schedules, which is reassuring partial evidence. However, the manuscript never trains with the proposed VP-ISSNR schedule itself. Since schedule transfer is only as good as the score approximation over the queried (x̂, csnr(γ^2)) pairs, a decisive experiment that trains with VP-ISSNR (or demonstrates schedule-ranking invariance across more than two training schedules) is needed to support the claim that VP-ISSNR is a state-of-the-art scheduling choice for few-step molecular generation.","section":"Section 4.1 and Appendix C"}],"minor_comments":[{"comment":"The displayed formula for γ^2(t) is inconsistent with Table A1 and with the boundary expressions in Eqs. (11) and (12): as written, the argument of the logarithm becomes negative for t close to tmax (e.g., with tmax=0.99). The intended schedule appears to be γ^2(t) = ((1-t)/t)^{2η} exp(2κ) with a truncation/renormalization to the interval [tmin, tmax]; please repair the formula so that a reader can implement it directly.","section":"Section 3.3, Eq. (10)"},{"comment":"The statement of an 'up to 30-fold' improvement in molecular generation is ambiguous; the authors should specify the comparison point, e.g., the NFE at which the best baseline reaches the same stability as VP-ISSNR at 4 or 8 NFEs.","section":"Section 5, Conclusion"},{"comment":"The curvature measure E[||(x(1)-x(0))-ẋ(t)||^2] is used to support the trajectory-straightness hypothesis, but the manuscript does not explain why this quantity controls the discretization error of the reverse ODE; a short derivation or a reference would make the argument easier to evaluate.","section":"Section 4.3"},{"comment":"The definition of the EDM non-uniform time grid uses σ_{i<N} and then sets σ_N = 0, which is standard but can be confusing; adding an explicit note that σ_N = 0 corresponds to the data endpoint would improve readability.","section":"Appendix A, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The algebraic error in the drift derivation (Appendix B.3) is serious and should be checked against the code before any decision. If the code uses the erroneous f, the experimental comparisons could change substantially after the correction. The per-NFE Bayesian tuning for images also needs clarification on the selection protocol. The molecular results, especially the fixed-parameter VP-ISSNR, are the strongest part of the paper, and the EDM-trained-model check in Figure A4 is a good step toward addressing schedule transfer."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth refereeing. The TV/SNR disentanglement in Eq. (8) is a simple algebraic reparameterization, but it is a useful one: it unifies a dozen existing schedules in Table A1 and gives you a clean language for asking what actually matters in the noise schedule. The SDE derivations in Appendix B are standard and correct. The empirical observation that constant-TV variants of VE schedules (SMLD, EDM-UT) improve few-step sampling is the most solid new finding. On QM9 the VP variants clearly beat their exploding-TV parents, and the effect is consistent across Euler, Heun, and DPM solvers. That alone is a publishable result.\n\nThe soft spot is exactly what the stress-test note says: the headline VP-ISSNR numbers come from a model trained with DDPM-cos (ν=1) and τ(t)=1, not with the proposed schedule. The authors condition on scaled SNR and show this makes the model compatible across schedules, and Appendix A4 does train an EDM-schedule model and shows VP-ISSNR still wins. That partially mitigates the concern. But it is not the same as training with VP-ISSNR itself, and the missing experiment leaves the central claim entangled with the training schedule. A retrained VP-ISSNR model is the decisive check.\n\nImage results are weaker support. EDM remains the best on CIFAR-10 and FFHQ unless η and κ are Bayesian-tuned per NFE against FID, which is a tuned result. The fixed-parameter variant is merely competitive. No error bars anywhere, so the differences at low NFE are hard to judge. On novelty, the inverse sigmoid SNR is close in spirit to the sigmoid schedules in Lin et al. (2024), though the constant-TV twist and the trajectory-curvature analysis are new.\n\nThis is a serious paper and the authors are honest about the evaluation limits. I would send it to peer review, and the main referee request should be the retraining experiment plus error bars on the key curves. If retraining confirms the gains, this is a solid contribution; if not, the constant-TV finding still stands alone.","headline":"A clean scheduling reformulation with a genuinely useful constant-TV insight, but the headline molecular gains rest on a model trained with a different schedule and need a retraining check.","tokens_in":25266,"tokens_out":1776,"would_cite":true,"duration_ms":20955,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that diffusion-model sampling can be accelerated substantially by separating total variance from signal-to-noise ratio, that exploding-variance schedules are usually dominated by constant-variance variants with the same…","keywords":["diffusion models","noise schedule","total variance","signal-to-noise ratio","inverse sigmoid SNR","flow matching","few-step sampling","molecular structure generation"],"falsifier":"Train a fresh score network from scratch under the VP-ISSNR schedule itself, with matched architecture and compute, and compare stability and FID at the same NFEs against the paper's numbers obtained by reusing a cosine-trained model. The appendix already hints that the cosine-trained model beats an EDM-trained model even when the EDM schedule is used for sampling; if consistent retraining removes or reverses the VP-ISSNR advantage, the schedule-specific claims would collapse. A cleaner test is to train and evaluate every candidate schedule under both matched and cross-schedule conditions and check whether the rank order of schedules survives matched training.","tokens_in":24277,"feed_emoji":"🧪","tokens_out":10606,"duration_ms":94379,"temperature":0.7,"pith_summary":"This paper argues that the noise schedule of a diffusion model should be designed with two independent dials: the total variance (TV) of the perturbed signal and its signal-to-noise ratio (SNR). Working in that disentangled framework, the authors show that common schedules with exponentially growing total variance, such as SMLD and the uniform-time variant of EDM, are improved by switching to a constant TV while keeping the same SNR schedule. They then propose VP-ISSNR, a constant-TV schedule whose SNR decays like the exponential of an inverse sigmoid function, which generalizes optimal-transport flow matching. On QM9 molecular structure generation this schedule yields stable molecules with as few as four solver steps and state-of-the-art stability at 32 to 64 steps, with similar or better few-step image quality on CIFAR-10 and FFHQ. If the claim holds, sampling cost in diffusion models can be cut sharply without retraining, and schedule design becomes an axis that composes with better solvers.","feed_headline":"Stable molecules in four steps with a constant-variance schedule","feed_subtitle":"Separating total variance from signal-to-noise ratio lets the VP-ISSNR schedule cut diffusion sampling cost dramatically","key_machinery":"The central object is the TV/SNR parameterization of the perturbation kernel, which rewrites the forward noising distribution in terms of TV $\\tau^2=a^2+b^2$ and SNR $\\gamma=a/b$, with the corresponding SDE drift and diffusion coefficients derived from these two functions. This separation allows a schedule's noise magnitude and its signal retention to be changed independently. The proposed VP-ISSNR schedule fixes $\\tau^2=1$ and sets $\\gamma^2(t)=((1-t)/t)^{2\\eta}e^{2\\kappa}$, the exponential of an inverse sigmoid in $\\log(1/t-1)$, which produces rapid SNR decay near both ends of the diffusion interval and reduces to the OTFM SNR when $\\eta=1$, $\\kappa=0$. The second load-bearing piece is the SNR-conditioned score network: the model is trained with constant TV and conditioned on a scaled, log-linearized SNR, $c_{\\text{snr}}(\\gamma^2)=\\omega\\log\\gamma^2+\\xi$, which lets one trained network be sampled under many schedules and makes the paper's cross-schedule comparisons possible.","core_discovery":"The paper's central claim is that the total variance (TV), $\\tau^2(t) = a^2(t)+b^2(t)$, and the signal-to-noise ratio (SNR), $\\gamma(t)=a(t)/b(t)$, of a diffusion perturbation kernel are separate design choices, and that treating them as such improves fast sampling. Reformulating the kernel as $p(x(t)|x(0))=\\mathcal{N}\\big(\\sqrt{\\tau^2\\gamma^2/(1+\\gamma^2)}\\;x(0),\\; \\tau^2/(1+\\gamma^2)\\;I\\big)$, the authors show that schedules whose TV explodes exponentially, such as SMLD and EDM with a uniform time grid, are dominated at equal SNR by their constant-TV (VP) counterparts in both molecular and image generation. They propose the VP-ISSNR schedule, $\\tau^2(t)=1$ with $\\gamma^2(t)=((1-t)/t)^{2\\eta}e^{2\\kappa}$, an inverse-sigmoid SNR law that contains optimal-transport flow matching as a special case. With this schedule, a model trained once under a cosine SNR schedule generates stable QM9 molecules in four Euler steps, reaches roughly 87 percent stability at 128 steps, and exceeds 95 percent with stochastic sampling at 64 steps; on images it matches or improves the highly optimized EDM sampler at low step counts when the two schedule parameters are scaled or tuned. The paper also offers a mechanistic explanation from a toy mixture-of-delta experiment: good schedules keep ODE trajectories straight near the data manifold and grow the marginal support early.","pith_inferences":["If the SNR-conditioned transfer is general, the TV/SNR split implies that existing pretrained diffusion checkpoints could be re-scheduled at inference time, per dataset or per NFE, without retraining; the paper demonstrates this capability for its own experiments but does not develop it into a general recipe for arbitrary checkpoints.","The inverse-sigmoid SNR allocates many integration steps near both $t\\approx 0$ and $t\\approx 1$, not just near the data manifold; a testable extension is that the high-noise end matters for high-frequency detail, which could be checked by ablating the offset parameter $\\kappa$ while holding the total number of steps fixed.","The fact that VP-OTFM matches OTFM on molecules but underperforms on images hints that the optimal TV schedule is task-dependent; searching over $\\tau(t)$ jointly with $\\gamma(t)$ using the curvature and support diagnostics is a natural next step that the paper leaves open.","The curvature and support hypothesis could be turned into a training-free schedule optimizer: parametrize $\\tau$ and $\\gamma$, and minimize integrated trajectory curvature subject to a support-growth constraint on a small calibration set; the paper supplies the diagnostics but not the optimization."],"forward_implications":["Schedules with exponentially exploding total variance (SMLD and uniform-time EDM) can be converted to constant-TV variants with the same SNR curve, improving or matching stability and FID at every tested number of function evaluations.","The VP-ISSNR schedule produces stable QM9 molecules with a first-order Euler ODE in just four NFEs, about 74 percent stability at eight NFEs, and nearly 87 percent at 128 NFEs; with the reverse SDE it reaches 93.16 percent at 32 NFEs and 95.82 percent at 64 NFEs, which the authors report as state of the art for those budgets.","The schedule benefit persists across Euler, Heun, DPM, and RK45 solvers, so noise-schedule optimization is orthogonal to solver improvements and can be combined with them.","On image datasets (CIFAR-10, FFHQ, AFHQ, ImageNet), VP-ISSNR with an NFE-scaled or tuned schedule parameter matches or improves EDM and OTFM FID in the low-NFE regime while reusing the same pretrained EDM score networks.","The toy trajectory analysis suggests a concrete design criterion for fast schedules: low ODE curvature near the data manifold at $t\\approx 0$ and marginal support that saturates before $t=1$.","Because the molecular model is trained once under a cosine SNR schedule and then sampled under many schedules, the reported gains do not require retraining per schedule, provided the SNR-conditioned score transfer holds."],"supporting_citations":[{"why":"Provides the EDM framework, the non-uniform time grid, the pretrained image models, and the main image-generation baseline that VP variants and VP-ISSNR are compared against.","marker":"Karras et al. [2022]"},{"why":"Defines optimal-transport flow matching, whose SNR schedule the proposed ISSNR schedule generalizes and which serves as a principal baseline in both molecule and image experiments.","marker":"Lipman et al. [2023]"},{"why":"Introduces DDPM and the original variance-preserving schedule, providing the linear and cosine VP baselines used in the comparisons.","marker":"Ho et al. [2020]"},{"why":"Supplies the score-based SDE framework and the reverse-SDE/ODE formulation used for sampling with the various schedules.","marker":"Song et al. [2021a]"},{"why":"Introduces SMLD and the variance-exploding schedule whose VP variant is a core demonstration of the paper's constant-TV claim.","marker":"Song and Ermon [2019]"},{"why":"Defines the SNR as $a(t)/b(t)$ and provides the variational-diffusion perspective that the TV/SNR disentanglement builds on.","marker":"Kingma et al. [2021]"},{"why":"Supplies the molecular relaxation model architecture, the QM9 training split, and the experimental setup used for molecular structure generation.","marker":"Kahouli et al. [2024]"},{"why":"Introduces the cosine noise schedule, including the $\\nu$ parameterization used to train the molecular score network under DDPM-cos.","marker":"Nichol and Dhariwal [2021]"},{"why":"Provides the trajectory-curvature measure $E[\\|(x(1)-x(0))-\\dot{x}(t)\\|^2]$ used in the numerical analysis connecting low curvature to fast sampling.","marker":"Liu et al. [2023b]"}],"fun_headline_variants":["Constant variance meets SNR tuning for fast diffusion","Splitting variance and SNR improves diffusion sampling","TV/SNR separation enables four-step molecular generation","Rethink diffusion noise: constant TV, sharper sampling"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that a score network trained under one SNR schedule (a cosine schedule for molecules, pretrained EDM models for images) can be sampled under a different SNR schedule without retraining, so that the observed differences in stability and FID are caused by the schedule rather than by the training run.","fun_headline_variants_meta":{"raw":{"variants":["Constant variance meets SNR tuning for fast diffusion","Splitting variance and SNR improves diffusion sampling","TV/SNR separation enables four-step molecular generation","Rethink diffusion noise: constant TV, sharper sampling"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000264,"raw_usage":{"total_tokens":1657,"prompt_tokens":1053,"completion_tokens":604,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":543}},"tokens_in":669,"tokens_out":604,"duration_ms":7259,"temperature":1.0,"reasoning_tokens":543,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T04:30:59.508098+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a fresh score network from scratch under the VP-ISSNR schedule itself, with matched architecture and compute, and compare stability and FID at the same NFEs against the paper's numbers obtained by reusing a cosine-trained model. The appendix already hints that the cosine-trained model beats an EDM-trained model even when the EDM schedule is used for sampling; if consistent retraining removes or reverses the VP-ISSNR advantage, the schedule-specific claims would collapse. A cleaner test is to train and evaluate every candidate schedule under both matched and cross-schedule conditions and check whether the rank order of schedules survives matched training.","supporting_citations":[{"cited_title":"Elucidating the design space of diffusion-based generative models","cited_arxiv_id":null,"evidence_quote":"Provides the EDM framework, the non-uniform time grid, the pretrained image models, and the main image-generation baseline that VP variants and VP-ISSNR are compared against."},{"cited_title":"Molecular relaxation by reverse diffusion with time step prediction","cited_arxiv_id":null,"evidence_quote":"Supplies the molecular relaxation model architecture, the QM9 training split, and the experimental setup used for molecular structure generation."}],"review_version":1}