Pith. sign in

REVIEW 3 major objections 4 minor 35 references

This paper claims that, for a Gaussian source with a given covariance, the reverse-sampling error of a diffusion model is governed by a noise schedule that follows a tangent law, with a single optimal coefficient set by the trace ratio of t

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 09:28 UTC pith:TIPMJQ6V

load-bearing objection Clean analytic core — closed-form Gaussian DDIM KL, tangent-law schedule, derived γ* — with a real but honestly flagged gap between the Theorem 1 lower bound and the Section IV deterministic-sampler optimization. the 3 major comments →

arxiv 2601.13602 v3 pith:TIPMJQ6V submitted 2026-01-20 cs.IT cs.LGmath.IT

A Gaussian Perspective for Distributional Discrepancy in Generative Diffusion Models

Qiang Sun , H. Vincent Poor , Wenyi Zhang This is my paper
classification cs.IT cs.LGmath.IT MSC 94A1749K1060J60
keywords diffusion modelsnoise scheduleKL divergencetangent lawGaussian sourceEuler-Maclaurin expansiontime discretizationdeterministic reverse sampling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to show that the sampling error of a diffusion model can be understood from first principles by replacing the data distribution with a Gaussian having the same covariance. For that Gaussian, the reverse-sampling KL divergence has a closed form whose leading term, at step count N, is a functional of the noise schedule. Minimizing that functional yields a noise schedule following a tangent law, with a single global coefficient γ* = sqrt(tr Σ / tr Σ⁻¹). The paper further argues the Gaussian case is a strict lower bound for general distributions, and uses the closed-form KL as a cheap metric to pick time-discretization steps that outperform existing heuristics at low function-evaluation budgets.

Core claim

On its own terms, the paper establishes a closed-form description of how a Gaussian source is transformed by deterministic reverse diffusion sampling. Treating the source as N(0, Σ), it derives the exact eigenvalue map per sampling step, writes the KL divergence between source and reconstructed distribution in closed form, and shows by Euler–Maclaurin expansion that this KL decays as O(1/N²) with a leading coefficient fixed by the schedule. Calculus of variations over that coefficient yields the tangent-law schedule η_ℓ(t) = √μ_ℓ tan(πt/2) for each covariance mode, and a unified schedule with γ* = √(tr Σ / tr Σ⁻¹). The paper then uses the same KL formula as a retraining-free score to rank ti

What carries the argument

The central object is the Gaussian reference process: a multivariate Gaussian source with the same covariance as the data. Because the optimal noise predictor is linear in this setting, each reverse step is a diagonal linear map in the eigenbasis, making the eigenvalue evolution explicit. The KL divergence between the source and reconstructed Gaussian is then computed in closed form, and its leading 1/N² term is written as the functional J_ℓ[η] = ∫ η_dot²/(μ_ℓ+η²)² dt. The change of variables Q_ℓ = (1/√μ_ℓ) arctan(η/√μ_ℓ) turns J_ℓ into ∫ Q_dot² dt, whose Euler–Lagrange solution is a straight line — hence the tangent law. This Q-transformation is the mechanism that carries the variational ar

Load-bearing premise

The practical transfer rests on the belief that a Gaussian source with a synthetic power-law spectrum captures the behavior of real image distributions and of already-trained score networks; if that proxy is not representative, the recommended schedules need not minimize the true reverse-sampling discrepancy.

What would settle it

Take a non-Gaussian source with the same covariance (e.g., a mixture of two well-separated Gaussians), run deterministic reverse sampling with the tangent-law schedule and with uniform-λ at the same N, and compute the true KL against the source. If the tangent-law schedule's KL is not smaller than uniform-λ's, the lower-bound-to-minimizer transfer fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The closed-form KL gives a training-free metric for comparing time discretizations of any pretrained diffusion model, with no search or retraining needed.
  • The tangent-law schedule is the unique minimizer of the per-mode asymptotic error, so any schedule that achieves the best Gaussian KL must approximate it.
  • The explicit γ* = sqrt(tr Σ / tr Σ⁻¹) ties the schedule to the covariance spectrum: the optimal unified schedule depends on the data through the trace ratio of the covariance and its inverse.
  • Sampling error decays as O(1/N²) under smooth schedules, so the gains from schedule choice concentrate at small N, i.e., when the number of function evaluations is tightly budgeted.
  • The power-uniform discretization with ρ≈1.5–2.0 concentrates steps in the high-noise region, matching the rapid η decay of the tangent law, and yields the reported FID improvements.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the Gaussian lower bound is representative for the leading term, schedule design for real diffusion models could be reduced to spectral estimation: measure or fit the source covariance's eigenvalue decay and read off the schedule; the paper only demonstrates this with a synthetic power-law spectrum, leaving a direct spectral-estimation pipeline as a natural extension.
  • The same variational argument likely extends to variance-exploding schedules with finite terminal noise; the paper shows the VP coefficient is recovered asymptotically, but a fully closed-form finite-σ_max variant could be derived from the arctan endpoint in the appendix.
  • The tangent-law schedule rescales time differently for each eigenvalue mode, suggesting that an implementation with per-mode time warping, rather than a single global warping, could push the Gaussian KL even lower.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper develops an analytical, Gaussian-source treatment of distributional discrepancy in diffusion models. It derives the closed-form output covariance of deterministic reverse sampling under the optimal linear estimator (Lemma 1), the resulting Gaussian KL divergence (Eq. 26), and its O(1/N^2) leading term via the Euler-Maclaurin expansion (Lemma 2). Minimizing this leading term yields a tangent-law noise schedule (Theorem 2) and, for a shared scalar scale, the closed-form optimizer gamma* = sqrt(tr Sigma_x / tr Sigma_x^{-1}) (Theorem 3). A separate extremal claim (Theorem 1) asserts that the Gaussian source provides a lower bound on the stochastic path KL for general sources with the same covariance. The paper then uses the closed-form KL to select a power-uniform time discretization, and reports FID improvements on CIFAR-10 and FFHQ-64 with pretrained EDM and VP-SDE models. The Gaussian derivations (Lemmas 1-2, Theorems 2-3) are internally coherent; the main unresolved issues concern the proof of Theorem 1 and the transfer from the stochastic path-KL lower bound to the deterministic sampler and to real pretrained models.

Significance. If the Gaussian lower bound and the transfer to deterministic sampling were fully established, the closed-form KL, the tangent law, and the gamma* formula would provide a valuable first-principles reference for schedule design and step selection. The analytical results are elegant and falsifiable: for a Gaussian source, Eqs. (24), (30), and (41) give explicit, testable predictions, and Appendix D-B provides synthetic validation. The experiments with pretrained models are extensive and show consistent advantages over uniform-lambda, though the comparison with Xue'24 is mixed in some regimes. The significance of the paper therefore rests on two pillars: the rigorous Gaussian optimization, which is largely sound, and the Gaussian-proxy transfer to real data, which is currently heuristic rather than proven.

major comments (3)
  1. [Section III, Theorem 1 (Eqs. (13), (20), and Appendix E, Eq. (98))] The proof that Delta >= 0 is incomplete. Eq. (13) writes Delta = Delta_init + sum_j (1/4) h_{t_j}^2 delta_j^2 [ E || -sigma_{t_j} nabla^2 log q_{t_j} ||_F^2 - || sigma_{t_j} Sigma_{t_j}^{-1} ||_F^2 ] + R(delta). The bracket is shown nonnegative, but R(delta)=O(delta^2) by Eq. (98), and the main term is also O(delta^2) since h_{t_j}=O(1). Thus 'sufficiently small step size' does not justify dropping R(delta) without a bound on its sign or a proof that the nonnegative term dominates it. In the Gaussian case the bracket vanishes and Delta is exactly the remaining R(delta), even though Delta=0. Because Theorem 1 is the stated motivation for the entire Gaussian analysis, this gap is load-bearing. Please provide a remainder bound with controlled sign, or explicitly add a condition under which the lower bound is valid.
  2. [Sections III and IV (transfer from path-KL bound to Eq. (26))] Theorem 1 concerns the stochastic reverse-time SDE path measures Q and P. Section IV immediately switches to the deterministic sampler (10) and uses the marginal KL (26). No theorem connects the stochastic path-KL lower bound to the deterministic sampler's marginal KL. The chain D_KL(Q||P) >= D_KL(Q_G||P_G) >= D_KL(q_{0,G}||p_{0,G}) has the wrong orientation to conclude D_KL(q0||p0) >= D_KL(q_{0,G}||p_{0,G}) for deterministic sampling: a lower bound for the stochastic path measure does not imply that minimizing the Gaussian marginal KL controls the deterministic marginal KL for a general q0. The correct reading appears to be that Theorems 2-3 are an exact Gaussian-source optimization for the deterministic sampler, and the use of (26) for real data is a proxy. This distinction should be made explicit, and preferably a deterministic-sampler analogue of Theorem 1 should be derived.
  3. [Section IV-C, Appendix F, and Table I] The recommended power-uniform exponents rho=1.5 and 2.0 are selected by evaluating the closed-form KL (26) on a synthetic shifted power-law covariance spectrum (Appendix F, with parameters C, i0, p, epsilon). This proxy is not validated against measured CIFAR-10 or FFHQ-64 covariance spectra. In addition, the KL (26) assumes the optimal linear estimator and the first-order update (10), whereas the experiments use pretrained score networks and the higher-order samplers DPM-Solver++ and UniPC. The transfer from a synthetic-spectrum KL ranking to FID improvements is therefore supported only empirically, not by the theoretical framework. Please report sensitivity of the rho selection to the spectrum parameters, compare with an empirical spectrum estimate from the datasets, and ideally test the selected schedule with a first-order sampler matching (10) to isolate the mechanism.
minor comments (4)
  1. [Lemma 1 and Appendix A] Lemma 1 states the initialization is N(0,I), but Appendix A and Eq. (52) use N(0, sigma_N^2 I). This is harmless for VP but not for VE, where sigma_N = sigma_max. Please correct the statement.
  2. [Section V, Table I (VP-SDE / UniPC rows)] The text says the strategy is 'consistently superior' to the search-based baseline, but Table I shows several NFEs (5-7) where Xue'24 outperforms both rho=1.5 and rho=2.0, e.g., NFE=5 FID 21.33 vs 22.24/28.49. The claim should be qualified.
  3. [Section IV-B and IV-C] The tangent-law schedule is derived for uniform time discretization, while the practical experiments use the power-uniform step rule (45). The relation between the schedule-optimization result and the step-selection result should be clarified, as they are different objects.
  4. [Abstract and Conclusion] The tangent-law schedule from Theorems 2-3 is not evaluated on real image data; the real-data experiments only test power-uniform time discretization of fixed pretrained schedules. The wording should not imply that the tangent law itself was validated on real data.

Circularity Check

0 steps flagged

No significant circularity: the Gaussian KL derivation is self-contained and the empirical schedules are not fitted to their own predictions.

full rationale

The paper's central chain is a direct computation, not an input-output tautology. Lemma 1 derives the Gaussian output covariance from the affine reverse update (10) and the closed-form posterior estimator (22); Lemma 2 obtains the O(1/N^2) KL coefficient by Euler-Maclaurin expansion; Theorem 2 solves the variational problem for J_l[η] exactly via the change of variables Q_l = arctan(η/√μ_l)/√μ_l, so the tangent law is the explicit Euler-Lagrange solution, not a pre-supposed ansatz; Theorem 3 minimizes a strictly convex closed-form function and derives γ* = sqrt(tr Σ_x / tr Σ_x^{-1}). The covariance eigenvalues are inputs, not fitted to the FID results, and the empirical ρ=1.5/2.0 choice is selected on a synthetic spectrum and then tested externally on CIFAR-10/FFHQ, so no fitted parameter is relabeled as a prediction. The only overlapping-author citation is Poor's textbook for the standard Gaussian conditional expectation formula, which is not load-bearing. Theorem 1's stochastic-path lower bound does not by itself justify the deterministic/non-Gaussian transfer, but that is an unproved modeling step rather than a circular reduction; no quoted equation reduces to its own input.

Axiom & Free-Parameter Ledger

2 free parameters · 6 axioms · 0 invented entities

The central derivation requires Gaussianity, an optimal estimator, and smoothness assumptions. Nothing fundamentally new is invented; the practical transfer to real data rests on a synthetic covariance proxy and hand-chosen ρ.

free parameters (2)
  • Power-law covariance spectrum parameters (C, i0, p, ε) = C calibrated to μ1=μmax; ε tuned to μk≥μmin; p and i0 left unspecified
    Used in Section IV-C and Appendix F to synthesize the covariance spectrum that mimics real image data; choice affects the ρ selection and KL evaluation.
  • Power-uniform discretization exponent ρ = 1.5 and 2.0
    Selected from Figure 1 based on the surrogate KL evaluation; no closed-form derivation or per-dataset optimization is provided.
axioms (6)
  • domain assumption Source q0 is multivariate Gaussian with a given covariance Σ_x and reverse sampling uses the optimal linear posterior estimator
    The closed-form KL analysis in Section IV and Lemma 1 rests entirely on this; real image data are not Gaussian.
  • domain assumption The score network is optimal, i.e., ε*_θ(x_t,t)=E[ε|x_t]=-σ_t ∇ log q_t(x_t)
    Theorem 1 and Lemma 1 assume the optimal estimator; pretrained networks are imperfect.
  • domain assumption Regularity conditions for Girsanov/Ito arguments: bounded Lipschitz SDE coefficients, C² in x and C¹ in t network, bounded moments, Novikov condition
    Assumed in Appendix E (Conditions 1-3) to justify the path-measure KL expansion; not verified for real models.
  • domain assumption α(t) and σ(t) are sufficiently smooth for the Euler-Maclaurin expansion and for the tangent-law boundary value problem
    Lemma 2 and Theorem 2 require smoothness; the tangent schedule has η→∞ at t=1, so the boundary condition is interpreted as a limit.
  • ad hoc to paper A synthetic shifted power-law eigenvalue spectrum is representative of CIFAR-10 and FFHQ-64 covariance decay
    Section IV-C and Appendix F; the actual data covariance is never measured, so the transfer of ρ=1.5/2.0 to real datasets is assumed.
  • domain assumption Minimizing the Gaussian lower bound also improves the general distribution mismatch
    Theorem 1 gives only a lower bound; the paper assumes the Gaussian proxy is representative for practical optimization, which is a leap beyond the proved inequality.

pith-pipeline@v1.3.0-alltime-deepseek · 23042 in / 18248 out tokens · 193569 ms · 2026-08-03T09:28:40.636549+00:00 · methodology

0 comments
read the original abstract

This paper introduces an analytical approach to quantifying and optimizing the distributional discrepancy in generative diffusion models. For a multivariate Gaussian source, we explicitly derive the closed-form evolution trajectory and the resulting Kullback-Leibler (KL) divergence between the distributions of the source data and the reversely sampled data. Asymptotic analysis via the Euler-Maclaurin expansion characterizes the convergence behavior of this KL divergence, extracting its dominant term as an explicit functional of the noise schedule. Minimizing this dominant term via the calculus of variations yields a noise schedule described by a tangent law, inherently determined by the source covariance spectrum. We further prove that the Gaussian source exhibits an extremal property for the KL divergence among general source distributions with a given covariance. We also utilize the analytical KL divergence as a principled metric to identify efficient time discretization strategies for pretrained diffusion models, and demonstrate via experiments over diverse datasets that the identified strategies consistently outperform established baselines, particularly under constrained function evaluation budgets.

Figures

Figures reproduced from arXiv: 2601.13602 by H. Vincent Poor, Qiang Sun, Wenyi Zhang.

Figure 1
Figure 1. Figure 1: KL divergence versus steps N for the time discretization strategy (45) under (a) VP and (b) VE settings. C. Time Discretization Strategy Evaluation In this subsection, we demonstrate another more direct application of the KL divergence-based analysis. Given that training a new DM is prohibitively computationally expensive for large datasets, in practice, users are typically constrained to pretrained models… view at source ↗
Figure 2
Figure 2. Figure 2: Performance analysis of the parameterized tangent law schedule in Gaussian settings. We report the KL [PITH_FULL_IMAGE:figures/full_fig_p022_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visualization of the ηi = e −λi under different time discretization strategies. b) Synthetic Source Distribution: Formally, the eigenvalues {µℓ} k ℓ=1 of the source covariance matrix are generated using a shifted power-law model. This formulation is designed to simulate the heavy-tailed eigenvalue spectra: µℓ = C · (ℓ + i0) −p + ε, ℓ = 1, . . . , k, (112) where p > 0 governs the decay rate, i0 > 0 smooths … view at source ↗
Figure 4
Figure 4. Figure 4: EDM checkpoint with DPM-Solver++ on CIFAR-10 (NFEs=6). [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: EDM checkpoint with DPM-Solver++ on CIFAR-10 (NFEs=12). [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: EDM checkpoint with DPM-Solver++ on CIFAR-10 (NFEs=15). [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: VP-SDE checkpoint with UniPC on CIFAR-10 (NFEs=20). [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: VP-SDE checkpoint with UniPC on CIFAR-10 (NFEs=25). [PITH_FULL_IMAGE:figures/full_fig_p028_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: VP-SDE checkpoint with UniPC on CIFAR-10 (NFEs=50). [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: EDM checkpoint with DPM-Solver++ on FFHQ-64 (NFEs=8). [PITH_FULL_IMAGE:figures/full_fig_p029_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: EDM checkpoint with DPM-Solver++ on FFHQ-64 (NFEs=10). [PITH_FULL_IMAGE:figures/full_fig_p029_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: EDM checkpoint with DPM-Solver++ on FFHQ-64 (NFEs=12). [PITH_FULL_IMAGE:figures/full_fig_p030_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 4 linked inside Pith

  1. [1]

    Deep unsupervised learning using nonequilibrium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” inProceedings of the 32nd International Conference on Machine Learning, vol. 37. PMLR, Jul. 2015, pp. 2256–2265

    2015

  2. [2]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” inAdvances in Neural Information Processing Systems, vol. 33, 2020, pp. 6840–6851

    2020

  3. [3]

    Score-based generative modeling through stochastic differential equations,

    Y . Song, J. Sohl-Dickstein, D. P. Kingma, A. Kumar, S. Ermon, and B. Poole, “Score-based generative modeling through stochastic differential equations,” inProceedings of the 9th International Conference on Learning Representations, 2021

    2021

  4. [4]

    Variational diffusion models,

    D. Kingma, T. Salimans, B. Poole, and J. Ho, “Variational diffusion models,” inAdvances in Neural Information Processing Systems, vol. 34, 2021, pp. 21 696–21 707

    2021

  5. [5]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, T. Aila, and S. Laine, “Elucidating the design space of diffusion-based generative models,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 26 565–26 577

    2022

  6. [6]

    Improved noise schedule for diffusion training,

    T. Hang, S. Gu, J. Bao, F. Wei, D. Chen, X. Geng, and B. Guo, “Improved noise schedule for diffusion training,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2025, pp. 4796–4806

    2025

  7. [7]

    Using Ornstein-Uhlenbeck process to understand denoising diffusion probabilistic model and its noise schedules,

    J. E. Santos and Y . T. Lin, “Using Ornstein-Uhlenbeck process to understand denoising diffusion probabilistic model and its noise schedules,” arXiv, 2023. [Online]. Available: https://arxiv.org/abs/2311.17673

    Pith/arXiv arXiv 2023

  8. [8]

    The cosine schedule is Fisher-Rao-optimal for masked discrete diffusion models,

    L. Zhang and S. Syed, “The cosine schedule is Fisher-Rao-optimal for masked discrete diffusion models,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2508.04884

    arXiv 2025

  9. [9]

    Denoising diffusion implicit models,

    J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” inProceedings of the 9th International Conference on Learning Representations, 2021

    2021

  10. [10]

    DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. LI, and J. Zhu, “DPM-Solver: A fast ODE solver for diffusion probabilistic model sampling in around 10 steps,” inAdvances in Neural Information Processing Systems, vol. 35, 2022, pp. 5775–5787

    2022

  11. [11]

    DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models,

    C. Lu, Y . Zhou, F. Bao, J. Chen, C. Li, and J. Zhu, “DPM-Solver++: Fast solver for guided sampling of diffusion probabilistic models,” Machine Intelligence Research, vol. 22, no. 4, pp. 730–751, 2025. February 10, 2026 DRAFT

    2025

  12. [12]

    UniPC: A unified predictor-corrector framework for fast sampling of diffusion models,

    W. Zhao, L. Bai, Y . Rao, J. Zhou, and J. Lu, “UniPC: A unified predictor-corrector framework for fast sampling of diffusion models,” in Advances in Neural Information Processing Systems, vol. 36, 2023, pp. 49 842–49 869

    2023

  13. [13]

    DPM-Solver-v3: Improved diffusion ODE solver with empirical model statistics,

    K. Zheng, C. Lu, J. Chen, and J. Zhu, “DPM-Solver-v3: Improved diffusion ODE solver with empirical model statistics,” inAdvances in Neural Information Processing Systems, vol. 36, 2023, pp. 55 502–55 542

    2023

  14. [14]

    Learning to schedule in diffusion probabilistic models,

    Y . Wang, X. Wang, A.-D. Dinh, B. Du, and C. Xu, “Learning to schedule in diffusion probabilistic models,” inProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2023, pp. 2478–2488

    2023

  15. [15]

    Autodiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration,

    L. Li, H. Li, X. Zheng, J. Wu, X. Xiao, R. Wang, M. Zheng, X. Pan, F. Chao, and R. Ji, “Autodiffusion: Training-free optimization of time steps and architectures for automated diffusion model acceleration,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), Oct. 2023, pp. 7105–7114

    2023

  16. [16]

    Align your steps: Optimizing sampling schedules in diffusion models,

    A. Sabour, S. Fidler, and K. Kreis, “Align your steps: Optimizing sampling schedules in diffusion models,” inProceedings of the 41st International Conference on Machine Learning, vol. 235. PMLR, Jul. 2024, pp. 42 947–42 975

    2024

  17. [17]

    Accelerating diffusion sampling with optimized time steps,

    S. Xue, Z. Liu, F. Chen, S. Zhang, T. Hu, E. Xie, and Z. Li, “Accelerating diffusion sampling with optimized time steps,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 8292–8301

    2024

  18. [18]

    Score-optimal diffusion schedules,

    C. Williams, A. Campbell, A. Doucet, and S. Syed, “Score-optimal diffusion schedules,” inAdvances in Neural Information Processing Systems, vol. 37, 2024, pp. 107 960–107 983

    2024

  19. [19]

    Variance-reduction guidance: Sampling trajectory optimization for diffusion models,

    S. Xu, Y . Liu, and A. W.-K. Kong, “Variance-reduction guidance: Sampling trajectory optimization for diffusion models,” inProceedings of the IEEE International Conference on Multimedia and Expo (ICME), 2025, pp. 1–6

    2025

  20. [20]

    Hierarchical schedule optimization for fast and robust diffusion model sampling,

    A. Zhu, R. Su, Q. Zhao, L. Feng, M. Shen, and S. He, “Hierarchical schedule optimization for fast and robust diffusion model sampling,” arXiv, 2025. [Online]. Available: https://arxiv.org/abs/2511.11688

    Pith/arXiv arXiv 2025

  21. [21]

    Zwillinger,CRC Standard Mathematical Tables and Formulae

    D. Zwillinger,CRC Standard Mathematical Tables and Formulae. Chapman and Hall/CRC, 2002

    2002

  22. [22]

    Learning multiple layers of features from tiny images,

    A. Krizhevsky, “Learning multiple layers of features from tiny images,” University of Toronto, Toronto, Ontario, Tech. Rep., 2009

    2009

  23. [23]

    A style-based generator architecture for generative adversarial networks,

    T. Karras, S. Laine, and T. Aila, “A style-based generator architecture for generative adversarial networks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 4401–4410

    2019

  24. [24]

    Reverse-time diffusion equation models,

    B. D. Anderson, “Reverse-time diffusion equation models,”Stochastic Processes and their Applications, vol. 12, no. 3, pp. 313–326, 1982

    1982

  25. [25]

    Tweedie’s formula and selection bias,

    B. Efron, “Tweedie’s formula and selection bias,”Journal of the American Statistical Association, vol. 106, no. 496, pp. 1602–1614, 2011

    2011

  26. [26]

    Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions,

    S. Chen, S. Chewi, J. Li, Y . Li, A. Salim, and A. R. Zhang, “Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions,”arXiv preprint arXiv:2209.11215, 2022

    Pith/arXiv arXiv 2022

  27. [27]

    Let us build bridges: Understanding and extending diffusion generative models,

    X. Liu, L. Wu, M. Ye, and Q. Liu, “Let us build bridges: Understanding and extending diffusion generative models,”arXiv preprint arXiv:2208.14699, 2022

    Pith/arXiv arXiv 2022

  28. [28]

    T. M. Cover,Elements of Information Theory. John Wiley & Sons, 1999

    1999

  29. [29]

    H. V . Poor,An Introduction to Signal Detection and Estimation. Springer Science & Business Media, 2013

    2013

  30. [30]

    Statistics of natural image categories,

    A. Torralba and A. Oliva, “Statistics of natural image categories,”Network: Computation in Neural Systems, vol. 14, no. 3, pp. 391–412, 2003

    2003

  31. [31]

    GANs trained by a two time-scale update rule converge to a local nash equilibrium,

    M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “GANs trained by a two time-scale update rule converge to a local nash equilibrium,” inAdvances in Neural Information Processing Systems, vol. 30, 2017, pp. 6626–6637

    2017

  32. [32]

    I. S. Gradshteyn and I. M. Ryzhik,Table of Integrals, Series, and Products. Academic Press, 2014

    2014

  33. [33]

    Flow matching for generative modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow matching for generative modeling,” inProceedings of the 11th International Conference on Learning Representations, 2023

    2023

  34. [34]

    Improved denoising diffusion probabilistic models,

    A. Q. Nichol and P. Dhariwal, “Improved denoising diffusion probabilistic models,” inProceedings of the 38th International Conference on Machine Learning, vol. 139. PMLR, Jul. 2021, pp. 8162–8171

    2021

  35. [35]

    Øksendal,Stochastic Differential Equations: An Introduction with Applications, 6th ed., ser

    B. Øksendal,Stochastic Differential Equations: An Introduction with Applications, 6th ed., ser. Universitext. Berlin, Heidelberg: Springer Berlin, Heidelberg, 2003. February 10, 2026 DRAFT APPENDIXA PROOF OFLEMMA1 This proof shows that, under the optimal estimator given by (7) and (22), the explicit closed-form evolution trajectory of the distributions ac...

    2003