Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Efficiently Access Diffusion Fisher: Within the Outer Product Span Space

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The diffusion Fisher matrix sits in the span of outer products of the initial data, so its trace and matrix-vector products can be read off with vector products instead of second-order auto-differentiation.

desk verdict Useful engineering with a correct core identity, but the DF-EA theory doesn't close; worth reviewing after revision. read the letter →

arxiv 2505.23264 v1 pith:UPEDNZS3 submitted 2025-05-29 cs.LG

Fangyikang Wang , Hubery Yin , Shaobin Zhuang , Huminhao Zhu , Yinan Li , Lei Qian , Chao Zhang , Hanbin Zhao
show 2 more authors
Hui Qian Chen Li
This is my paper
classification cs.LG MSC 68T07
keywords diffusionmodelsFisherinformationouter-productspantraceestimationadjointoptimizationoptimaltransportprobability-flowODElikelihoodevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Diffusion Fisher information—the negative Hessian of a diffusion model's marginal log-density—is usually obtained by differentiating the learned score network, a costly black-box step with no accuracy guarantee. This paper claims that the diffusion Fisher actually lies in the span of outer products of the initial data: it can be written as a schedule-dependent identity matrix minus a weighted covariance of data points. That structure turns two second-order operations, trace evaluation and matrix-vector multiplication, into vector-product computations with explicit error bounds. The payoff would be faster and more accurate likelihood evaluation and adjoint-guided sampling in large diffusion models, plus a numerical test for whether the probability-flow ODE is an optimal transport map.

What carries the argument

The carrying object is the outer-product span formulation of the diffusion Fisher: the Hessian of the log-marginal decomposes into an identity term plus a weighted covariance of the initial data, with weights determined by the diffusion kernel. Because $\operatorname{tr}(vv^\top)=\lVert v\rVert^2$ and $(vv^\top)\lambda=\langle v,\lambda\rangle v$, the formulation converts trace and matrix-vector access into norm and dot products. Two named mechanisms follow: DF-TM, a scalar network trained to match $\frac{1}{d}\mathbb{E}[\lVert x_0\rVert^2\mid x_t]$, and DF-EA, which substitutes the endpoint outer product $x_0x_0^\top$ for the weighted sum in adjoint ODEs. A third named mechanism is Corollary 1, which expresses optimal transport of the PF-ODE map as positive semidefiniteness of the normalized fundamental matrix of a matrix $B(t)$ built from the same weighted covariance.

What would settle it

At the discrete times used in 50-step adjoint sampling (guidance from step 15 to step 35), compute the exact weighted sum $\sum_i w_i y_i y_i^\top$ from the actual data distribution and measure its Hilbert–Schmidt distance to $x_0x_0^\top$; if that distance is comparable to or larger than the VJP approximation error, DF-EA's practical accuracy advantage is unsupported.

Watch

Extended reading notes

Core claim

At the paper's center is the identity $$F_t(x_t,t)=\frac{1}{\$sigma_t^{2}$}I-\frac{\$alpha_t^{2}$}{\$sigma_t^{4}$}\left[\int w(y)yy^\top dq_0(y)-\left(\int w(y)y\,dq_0(y)\right)\left(\int w(y)y\,dq_0(y)\right)^\top\right],$$ where $w(y)$ is the normalized Gaussian weight $\exp(-\lVert x_t-\alpha_t y\rVert^2/2\sigma_t^2)/\int \exp(-\lVert x_t-\alpha_t u\rVert^2/2\sigma_t^2)dq_0(u)$ and $q_0$ is the initial data distribution with finite second moment. In the empirical setting the integral becomes a weighted sum over the dataset. The paper proves that the first moment $\int w(y)y\,dq_0(y)$ is exactly what an optimally trained y-prediction score network estimates, so the only genuinely unknown term is the weighted second moment. DF-TM trains a scalar network to learn that term for trace access, while DF-EA replaces it with $x_0x_0^\top$ for gradient-free matrix-vector access; both come with error bounds. The same outer-product form yields a criterion for when the probability-flow ODE map is a Monge optimal transport map, and the paper's numerical test shows the criterion holds for Gaussian and affine initial data across common noise schedules but fails for a simple non-affine configuration.

Load-bearing premise

The load-bearing premise for DF-EA is that, at the finite times used in 50-step adjoint sampling, the weighted sum of data outer products can be replaced by the single endpoint outer product $x_0x_0^\top$, an approximation whose own error bound does not vanish as $t\to0$ or as score error goes to zero.

Editorial extensions

If this is right

  • DF-TM reduces per-sample trace evaluation from $O(d^2)$ to $O(d)$ and supplies a gradient-free likelihood ODE, with an error bound in terms of the trace-network and score-network errors.
  • DF-EA turns each adjoint Fisher application into scalar combinations of $\lambda_t$, $x_0$, and $\bar y_\theta(x_t,t)$, eliminating network backpropagation during guidance.
  • The paper's error analysis shows the DF-EA bound does not vanish as $t\to 0$ or as score error goes to zero, so the method's practical accuracy rests on the empirical comparisons reported at 50-step adjoint sampling.
  • For VE, VP, sub-VP, and EDM schedules, the numerical OT test returns zero asymmetry for single-Gaussian and affine initial data and clearly nonzero asymmetry for a three-point non-affine set, supporting the hypothesis that PF-ODE maps are optimal transport only for affine initial data.
  • Using DF-TM, likelihood and CLIP score form a trade-off curve across guidance scales in Stable Diffusion models, making per-sample negative log-likelihood a practical quality filter.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The authors leave implicit that the same weighted-covariance identity suggests a data-dependent correction to DF-EA, such as keeping the two or three largest-weight outer products, which could give a vanishing error bound at finite times.
  • Because the outer-product basis depends only on the data and noise schedule, random projections or sketching of $\sum_i w_i y_i y_i^\top$ could yield stochastic trace estimates with cheaper per-sample cost than training a dedicated trace network.
  • Corollary 1's fundamental-matrix test should transfer to any flow whose drift is an affine function of the score, such as flow matching or stochastic interpolants, not just diffusion ODEs.
  • The paper uses Euler integration and a fixed adjoint baseline; composing the same DF access with higher-order or exact-inversion ODE solvers could enlarge the measured speed-up.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper studies the diffusion Fisher information matrix F_t = −∇² log q_t for diffusion models. It derives exact formulas expressing F_t as 1/σ_t² I − α_t²/σ_t⁴ times a weighted covariance of the initial data, for both empirical (Dirac) and general initial distributions (Propositions 1 and 3). Based on these, it proposes DF-TM, a neural-network estimator for the trace, and DF-EA, an endpoint approximation for matrix-vector products, with claimed error bounds in Propositions 7 and 8. It also uses the outer-product form to numerically test whether PF-ODE maps are optimal transport (Corollary 1, Algorithm 2).

Significance. The main intellectual contribution is the exact outer-product span representation, which is a genuine structural insight that could enable cheaper access to second-order information in diffusion models. The trace-matching scheme is plausible, and the experiments demonstrate substantial practical speedups and competitive accuracy on likelihood evaluation and adjoint guidance. The optimal-transport verification is an interesting application of the formula. However, the paper advertises theoretical error bounds as the key advantage over VJP, and those bounds are currently not established as written: the proof of Proposition 7 contains an invalid inequality, and the Proposition 8 bound is non-vanishing and dimensionally inconsistent. The exact identities are correct and the empirical results are promising, so the paper has a solid core, but the theoretical claims need significant revision before the paper can be accepted.

major comments (3)
  1. [Appendix A.7 / Proposition 7] The proof bounds |∥Σwᵢyᵢ∥² − ∥ȳθ∥²| by (σ_t²/α_t²)δ₂², which is equivalent to asserting |∥a∥² − ∥b∥²| ≤ ∥a−b∥². This inequality is false in general; for example, a = 1 and b = 2 gives |1−4| = 3 while ∥a−b∥² = 1. A correct bound requires controlling ∥a∥+∥b∥, which is not done under the stated assumptions. Consequently, the error bound in Proposition 7, and with it the theoretical guarantee for DF-TM as stated, is not established.
  2. [Section 5, Eq. (20), Proposition 8, Appendix C.6] The DF-EA replacement of Σᵢ wᵢ yᵢ yᵢᵀ by x₀x₀ᵀ is uncontrolled in the regime where the method is deployed. As t→0, the weights wᵢ concentrate on the training point nearest to x_t/α_t, which for a generated x₀ is generally not the endpoint x₀; the approximation error therefore does not vanish. Appendix C.6 explicitly concedes that the bound is non-vanishing. In addition, the operator difference in Eq. (20) carries a prefactor α_t²/σ_t⁴, whereas the stated bound in Proposition 8 uses α_t²/σ_t³; the discrepancy needs to be resolved. These issues undermine the claim that DF-EA has a theoretical accuracy advantage over VJP.
  3. [Section 6 / Corollary 1 / Appendix A.10] The proof that the PF-ODE map is Monge only if the normalized fundamental matrix is s.p.d. assumes the flow Jacobian A(t) is symmetric so that a potential u exists by Poincaré's lemma. However, A(t) evolves by A′ = A B with B(t) symmetric; A(t) remains symmetric only if A commutes with B along the trajectory, which is not shown. Algorithm 3 also updates A_{i−1} = A_i + dt · A_iᵀB_i, while the derived ODE is A′ = A B, so the implementation and the theory appear inconsistent. The numerical OT conclusion therefore needs either a proof of symmetry or a reformulation as a necessary condition only.
minor comments (5)
  1. [Section 4, Eq. (17)] Equation (17) uses yθ, whereas the paper earlier defines ¯yθ; the notation should be unified throughout.
  2. [Table 1] The entries '0.07299%↓' and '0.06359%↓' are unclear; please state explicitly that these are relative time reductions compared with the VJP baseline.
  3. [Propositions 2 and 4] The main text says the optimal ¯yθ 'can accurately estimate' the weighted mean; it would be more precise to state that the minimizer of the denoising objective at fixed (x_t, t) equals the weighted mean, as proved in the appendix.
  4. [Table 3] The threshold for declaring 'Asym. OT' should be defined; the entries 0.00% are marked ✓ and 13–27% are marked ✗, but no acceptance criterion is given.
  5. [Abstract] The phrase 'the first numerical verification experiment' is strong given that the conclusion for non-affine data is only a hypothesis; consider softening to 'a numerical verification experiment'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; Proposition 3 is an exact identity, the DF-TM training target is an independent data moment, and the DF-EA endpoint replacement is a stated heuristic rather than a fitted prediction.

full rationale

The load-bearing derivation, Proposition 3 (and its Dirac analogue Proposition 1), is an exact analytic identity obtained by differentiating the log of the convolved marginal density under finite-second-moment assumptions; it does not assume the diffusion Fisher it claims to derive. Proposition 2 and Proposition 4 show that the optimal y-prediction of the standard denoising score-matching loss is exactly the conditional mean integral/sum appearing in the outer-product formula; this is a minimization of an independent regression target, not a fit to Fisher information. The DF-TM algorithm trains t_theta on the independent quantity ||x0||^2/d (Algorithm 1, Proposition 6) and then uses this learned component inside the exact trace identity, so no fitted parameter is renamed as a prediction. Proposition 7's error bound is a triangle-inequality estimate conditional on score and trace-network errors, not a restatement of the target. The DF-EA algorithm's replacement of the weighted outer-product sum by x0 x0^T (Eq. 20) is a heuristic justified by weight concentration near t = 0; Proposition 8 bounds the replacement error, and Appendix C.6 explicitly concedes that the bound is non-vanishing. That concession weakens the theoretical support for DF-EA as t -> 0 or as training error decreases, but this is a correctness/validity limitation, not circularity: the approximation is not fitted to the quantity it is claimed to predict, nor is it derived from a self-citation chain. The corollary on optimal transport is derived from external convex-analysis and optimal-transport theorems using the derived DF formula, and the numerical OT experiment is an independent check of prior results. Self-citations (e.g., Zhang et al. 2024a, Wang et al. 2024a, Zhu et al. 2024a/b) appear only as contextual or future-work references and are not load-bearing for the main claims. No circular step meeting the evidentiary standard could be identified.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

The derivations rest on standard Gaussian-convolution calculus plus two domain assumptions about the data distribution. The only ad hoc-to-paper item is the unproven symmetry of the PF-ODE Jacobian in the OT corollary. No new physical entities are postulated.

free parameters (1)
  • tθ(xt,t) trace network = learned scalar network; target is (1/d)Σ_i w_i ||y_i||²
    DF-TM requires training this network on data (Algorithm 1). Its approximation error δ1 is assumed small in Proposition 7, but no bound is derived from network capacity or training.
assumptions (6)
  • domain assumption q0 has finite second moment (q0 ∈ P2(Rd))
    Used in Proposition 3 and its proof (Section 3, Eq 12) to justify the general outer-product integral form.
  • domain assumption Data are bounded: there exists D_y with ||y_i|| ≤ D_y
    Used in the Dirac setting (Eq 10) to justify interchange of sum and gradient in Lemma 2 and to bound errors in Prop 8.
  • standard math Leibniz rule permits interchange of derivative with sum/integral
    Invoked in Appendix A.1/A.3 for the Dirac and general proofs; requires the bounded-moment assumptions above.
  • domain assumption The trained y-prediction network equals the posterior mean E[x0|xt]
    Propositions 2 and 4 show this for the optimal network; the practical algorithms assume the learned εθ/yθ is close enough that this identification holds (used in Eq 17 and Eq 20).
  • ad hoc to paper The PF-ODE flow map Jacobian A(t)=∂T_{t,T}/∂x_t is symmetric so a potential u exists
    Corollary 1's proof invokes Poincaré's lemma to write the map as ∇u, but symmetry of A is never established; the paper's own asymmetry metric treats it as nontrivial (Appendix A.10, Section 6).
  • standard math Brenier's theorem and Santambrogio's Theorem 1.48 characterize OT maps by convex potentials
    Used in Appendix A.10 to connect Monge optimality to positive semi-definiteness of the Hessian of the potential.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficiently Access Diffusion Fisher: Within the Outer Product Span Space." pith.science (2026). https://pith.science/paper/UPEDNZS3

@misc{pith2026250523264,
  author       = {Pith},
  title        = {Pith review of: Efficiently Access Diffusion Fisher: Within the Outer Product Span Space},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UPEDNZS3}},
  note         = {Machine review of arXiv:2505.23264}
}
read the original abstract

Recent Diffusion models (DMs) advancements have explored incorporating the second-order diffusion Fisher information (DF), defined as the negative Hessian of log density, into various downstream tasks and theoretical analysis. However, current practices typically approximate the diffusion Fisher by applying auto-differentiation to the learned score network. This black-box method, though straightforward, lacks any accuracy guarantee and is time-consuming. In this paper, we show that the diffusion Fisher actually resides within a space spanned by the outer products of score and initial data. Based on the outer-product structure, we develop two efficient approximation algorithms to access the trace and matrix-vector multiplication of DF, respectively. These algorithms bypass the auto-differentiation operations with time-efficient vector-product calculations. Furthermore, we establish the approximation error bounds for the proposed algorithms. Experiments in likelihood evaluation and adjoint optimization demonstrate the superior accuracy and reduced computational cost of our proposed algorithms. Additionally, based on the novel outer-product formulation of DF, we design the first numerical verification experiment for the optimal transport property of the general PF-ODE deduced map.

Figures

Figures reproduced from arXiv: 2505.23264 by the authors.

Figure 1
Figure 1. (a) The training loss of DF-TM for SD-1.5 and SD￾2base. It demonstrates commendable convergence behavior. (b) The trade-off curve of NLL and Clip score of SD-1.5 and SD-2base across various guidance scales in [1.5, 2.5, ..., 12.5, 13.5] initial distribution and the noise schedule. Proposition 1. Defines vi(xt, t) as exp − |xt−αtyi| 2 2σ 2 t  ∈ R and wi(xt, t) as Pvi(xt,t) j vj (xt,t) ∈ R. If q0 takes the form as i… view at source ↗
Figure 2
Figure 2. Our DF-TM method facilitates the effective evaluation of the NLL of generated samples with varying seeds. It can be demonstrated that a lower NLL signifies a region of higher possibility, thereby consistently indicating superior image quality. where tr(·) denotes the trace of a matrix, which is defined to be the sum of elements on the diagonal. Log-Likelihood Evaluation via VJP The current tech￾nique is only capable… view at source ↗
Figure 3
Figure 3. Comparison between our DF method and the VJP method on adjoint guidance sampling across five objective scores: SAC/AVA aesthetic score, Pick-Score, clip loss, and Face ID loss. Notably, our DF consistently achieves superior scores with less time expenditure. unknown element in equation 16 is P i wi∥yi∥ 2 . Conse￾quently, we suggest learning this term using a scalar-valued network, as per the following training schem… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual comparison of DF-EA (Ours) and VJP in the adjoint improvement task on (left) SAC aesthetic score and (right) Pick-Score. DF-EA consistently generates images with better visual effects and reduced time expenditure. NLL evaluation method proves to be an effective …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

71 extracted references · 36 canonical work pages

  1. [1]

    S., Boffi, N

    Albergo, M. S., Boffi, N. M., and Vanden-Eijnden, E. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023

    arXiv 2023

  2. [2]

    Ordinary differential equations: an introduction to nonlinear analysis, volume 13

    Amann, H. Ordinary differential equations: an introduction to nonlinear analysis, volume 13. Walter de gruyter, 2011

    work page 2011

  3. [3]

    Anderson, B. D. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12 0 (3): 0 313--326, 1982

    work page 1982

  4. [4]

    Estimating the optimal covariance with imperfect mean in diffusion probabilistic models

    Bao, F., Li, C., Sun, J., Zhu, J., and Zhang, B. Estimating the optimal covariance with imperfect mean in diffusion probabilistic models. arXiv preprint arXiv:2206.07309, 2022 a

    arXiv 2022

  5. [5]

    Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models

    Bao, F., Li, C., Zhu, J., and Zhang, B. Analytic-dpm: an analytic estimate of the optimal reverse variance in diffusion probabilistic models. arXiv preprint arXiv:2201.06503, 2022 b

    arXiv 2022

  6. [6]

    Bounding the Test Log-Likelihood of Generative Models

    Bengio, Y., Yao, L., and Cho, K. Bounding the test log-likelihood of generative models. arXiv preprint arXiv:1311.6184, 2013

    work page Pith review arXiv 2013

  7. [7]

    Benton, J. et al. Nearly d -linear convergence bounds for diffusion models via stochastic localization. In The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  8. [8]

    Blasingame, Z. W. and Liu, C. Adjointdeis: Efficient gradients for diffusion models. arXiv preprint arXiv:2405.15020, 2024

    work page Pith review arXiv 2024

Show all 71 references
  1. [9]

    D \'e composition polaire et r \'e arrangement monotone des champs de vecteurs

    Brenier, Y. D \'e composition polaire et r \'e arrangement monotone des champs de vecteurs. CR Acad. Sci. Paris S \'e r. I Math. , 305: 0 805--808, 1987

    1987

  2. [10]

    Polar factorization and monotone rearrangement of vector-valued functions

    Brenier, Y. Polar factorization and monotone rearrangement of vector-valued functions. Communications on pure and applied mathematics, 44 0 (4): 0 375--417, 1991

    1991

  3. [11]

    T., Rubanova, Y., Bettencourt, J., and Duvenaud, D

    Chen, R. T., Rubanova, Y., Bettencourt, J., and Duvenaud, D. K. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018

    2018

  4. [12]

    Chen, Z., Li, H., Wang, F., Zhang, O., Xu, H., Jiang, X., Song, Z., and Wang, E. H. Rethinking the diffusion models for numerical tabular data imputation from the perspective of wasserstein gradient flow. arXiv preprint arXiv:2406.15762, 2024

    2024 arXiv

  5. [13]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, volume 34, pp.\ 8780--8794, 2021 a

    2021

  6. [14]

    and Nichol, A

    Dhariwal, P. and Nichol, A. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34: 0 8780--8794, 2021 b

    2021

  7. [15]

    Pectp: Parameter-efficient cross-task prompts for incremental vision transformer

    Feng, Q., Zhao, H., Zhang, C., Dong, J., Ding, H., Jiang, Y.-G., and Qian, H. Pectp: Parameter-efficient cross-task prompts for incremental vision transformer. arXiv preprint arXiv:2407.03813, 2024 a

    2024 arXiv

  8. [16]

    Lw2g: Learning whether to grow for prompt-based continual learning

    Feng, Q., Zhou, D., Zhao, H., Zhang, C., and Qian, H. Lw2g: Learning whether to grow for prompt-based continual learning. arXiv preprint arXiv:2409.18860, 2024 b

    2024 arXiv

  9. [17]

    Iap: Improving continual learning of vision-language models via instance-aware prompting

    Fu, H., Zhao, H., Dong, J., Zhang, C., and Qian, H. Iap: Improving continual learning of vision-language models via instance-aware prompting. arXiv preprint arXiv:2503.20612, 2025

    2025 arXiv

  10. [18]

    A., Gohberg, I., Goldberg, S., and Kaashoek, M

    Gohberg, I., Goldberg, S., Kaashoek, M. A., Gohberg, I., Goldberg, S., and Kaashoek, M. A. Hilbert-schmidt operators. Classes of Linear Operators Vol. I, pp.\ 138--147, 1990

    1990

  11. [19]

    and Salimans, T

    Ho, J. and Salimans, T. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  12. [20]

    Denoising diffusion probabilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33: 0 6840--6851, 2020

    2020

  13. [21]

    J., Norouzi, M., and Salimans, T

    Ho, J., Saharia, C., Chan, W., Fleet, D. J., Norouzi, M., and Salimans, T. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23 0 (47): 0 1--33, 2022

    2022

  14. [22]

    Kantorovich, L. V. Mathematical methods of organizing and planning production. Management science, 6 0 (4): 0 366--422, 1960

    1960

  15. [23]

    Elucidating the design space of diffusion-based generative models

    Karras, T., Aittala, M., Aila, T., and Laine, S. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022

    2022 arXiv

  16. [24]

    Understanding DDPM latent codes through optimal transport

    Khrulkov, V., Ryzhakov, G., Chertkov, A., and Oseledets, I. Understanding DDPM latent codes through optimal transport. In International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=6PIrhAx1j4i

    2023

  17. [25]

    Variational diffusion models

    Kingma, D., Salimans, T., Poole, B., and Ho, J. Variational diffusion models. In Ranzato, M., Beygelzimer, A., Dauphin, Y., Liang, P., and Vaughan, J. W. (eds.), Advances in Neural Information Processing Systems, volume 34, pp.\ 21696--21707. Curran Associates, Inc., 2021. URL...

    2021

  18. [26]

    Fundamentals of differential geometry, volume 191

    Lang, S. Fundamentals of differential geometry, volume 191. Springer Science & Business Media, 2012

    2012

  19. [27]

    and Santambrogio, F

    Lavenant, H. and Santambrogio, F. The flow map of the fokker--planck equation does not provide optimal transport. Applied Mathematics Letters, 133: 0 108225, 2022

    2022

  20. [28]

    Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll \'a r, P., and Zitnick, C. L. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pp....

    2014

  21. [29]

    T., Ben-Hamu, H., Nickel, M., and Le, M

    Lipman, Y., Chen, R. T., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    2022 arXiv

  22. [30]

    Pseudo numerical methods for diffusion models on manifolds

    Liu, L., Ren, Y., Lin, Z., and Zhao, Z. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022

    2022 arXiv

  23. [31]

    and Hutter, F

    Loshchilov, I. and Hutter, F. Decoupled weight decay regularization. In International Conference on Learning Representations (ICLR), 2019

    2019

  24. [32]

    Maximum likelihood training for score-based diffusion ODE s by high order denoising score matching

    Lu, C., Zheng, K., Bao, F., Chen, J., Li, C., and Zhu, J. Maximum likelihood training for score-based diffusion ODE s by high order denoising score matching. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., and Sabato, S. (eds.), Proceedings of the 39th Intern...

    2022

  25. [33]

    Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35: 0 5775--5787, 2022 b

    2022

  26. [34]

    Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models

    Lu, C., Zhou, Y., Bao, F., Chen, J., Li, C., and Zhu, J. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022 c

    2022 arXiv

  27. [35]

    Limit formulas for the normalized fundamental matrix of the northwest-corner truncation of markov chains: Matrix-infinite-product-form solutions of block-hessenberg markov chains

    Masuyama, H. Limit formulas for the normalized fundamental matrix of the northwest-corner truncation of markov chains: Matrix-infinite-product-form solutions of block-hessenberg markov chains. arXiv preprint arXiv:1603.07787, 2016

    2016 arXiv

  28. [36]

    M \'e moire sur la th \'e orie des d \'e blais et des remblais

    Monge, G. M \'e moire sur la th \'e orie des d \'e blais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pp.\ 666--704, 1781

  29. [37]

    Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., and Chen, M

    Nichol, A. Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., Mcgrew, B., Sutskever, I., and Chen, M. GLIDE : Towards photorealistic image generation and editing with text-guided diffusion models. In Proceedings of the 39th International Conference on Machine Learning, volu...

    2022

  30. [38]

    Osler, T. J. Leibniz rule for fractional derivatives generalized and an application to infinite series. SIAM Journal on Applied Mathematics, 18 0 (3): 0 658--674, 1970

    1970

  31. [39]

    H., Tan, V

    Pan, J., Liew, J. H., Tan, V. Y., Feng, J., and Yan, H. Adjointdpm: Adjoint sensitivity method for gradient backpropagation of diffusion probabilistic models. arXiv preprint arXiv:2307.10711, 2023 a

    2023 arXiv

  32. [40]

    H., Feng, J., and Tan, V

    Pan, J., Yan, H., Liew, J. H., Feng, J., and Tan, V. Y. Towards accurate guided diffusion sampling through symplectic adjoint method. arXiv preprint arXiv:2312.12030, 2023 b

    2023 arXiv

  33. [41]

    Adjoint sensitivity analysis and optimization of hysteretic dynamic systems with nonlinear viscous dampers

    Pollini, N., Lavan, O., and Amir, O. Adjoint sensitivity analysis and optimization of hysteretic dynamic systems with nonlinear viscous dampers. Structural and Multidisciplinary Optimization, 57: 0 2273--2289, 2018

    2018

  34. [42]

    Hierarchical text-conditional image generation with clip latents

    Ramesh, A., Dhariwal, P., Nichol, A., Chu, C., and Chen, M. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022

    2022 arXiv

  35. [43]

    Rockafellar, R. T. Convex analysis:(pms-28). 2015

    2015

  36. [44]

    High-resolution image synthesis with latent diffusion models

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., and Ommer, B. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 10684--10695, 2022

    2022

  37. [45]

    L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., Ho, J., Fleet, D

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E. L., Ghasemipour, K., Gontijo Lopes, R., Karagol Ayan, B., Salimans, T., Ho, J., Fleet, D. J., and Norouzi, M. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural ...

    2022

  38. [46]

    and Ho, J

    Salimans, T. and Ho, J. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022

    2022 arXiv

  39. [47]

    Optimal transport for applied mathematicians

    Santambrogio, F. Optimal transport for applied mathematicians. Birk \"a user, NY , 55 0 (58-63): 0 94, 2015

    2015

  40. [48]

    W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C. W., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., Schramowski, P., Kundurthy, S. R., Crowson, K., Schmidt, L., Kaczmarczyk, R., and Jitsev, J. LAION -5b: An open large-scale dataset for training n...

    2022

  41. [49]

    Modern analysis of hutchinson's trace estimator

    Skorski, M. Modern analysis of hutchinson's trace estimator. In 2021 55th CISS. IEEE, 2021

    2021

  42. [50]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Sohl-Dickstein, J., Weiss, E., Maheswaranathan, N., and Ganguli, S. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pp.\ 2256--2265. PMLR, 2015

    2015

  43. [51]

    and Lai, H

    Song, K. and Lai, H. Fisher information improved training-free conditional diffusion model. arXiv preprint arXiv:2404.18252, 2024

    2024 arXiv

  44. [52]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  45. [53]

    P., Kumar, A., Ermon, S., and Poole, B

    Song, Y., Sohl-Dickstein, J., Kingma, D. P., Kumar, A., Ermon, S., and Poole, B. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  46. [54]

    Maximum likelihood training of score-based diffusion models

    Song, Y., Durkan, C., Murray, I., and Ermon, S. Maximum likelihood training of score-based diffusion models. Advances in neural information processing systems, 34: 0 1415--1428, 2021

    2021

  47. [55]

    Theis, L., Oord, A. v. d., and Bethge, M. A note on the evaluation of generative models. arXiv preprint arXiv:1511.01844, 2015

    2015 arXiv

  48. [56]

    Texttoucher: Fine-grained text-to-touch generation

    Tu, J., Fu, H., Yang, F., Zhao, H., Zhang, C., and Qian, H. Texttoucher: Fine-grained text-to-touch generation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 7455--7463, 2025 a

    2025

  49. [57]

    Driveditfit: Fine-tuning diffusion transformers for autonomous driving data generation

    Tu, J., Ji, W., Zhao, H., Zhang, C., Zimmermann, R., and Qian, H. Driveditfit: Fine-tuning diffusion transformers for autonomous driving data generation. ACM Transactions on Multimedia Computing, Communications and Applications, 21 0 (3): 0 1--29, 2025 b

    2025

  50. [58]

    Villani, C. et al. Optimal transport: old and new, volume 338. Springer, 2009

    2009

  51. [59]

    Edict: Exact diffusion inversion via coupled transformations

    Wallace, B., Gokul, A., and Naik, N. Edict: Exact diffusion inversion via coupled transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 22532--22541, 2023

    2023

  52. [60]

    Belm: Bidirectional explicit linear multi-step sampler for exact inversion in diffusion models

    Wang, F., Yin, H., Dong, Y.-J., Zhu, H., Zhao, H., Qian, H., Li, C., et al. Belm: Bidirectional explicit linear multi-step sampler for exact inversion in diffusion models. Advances in Neural Information Processing Systems, 37: 0 46118--46159, 2024 a

    2024

  53. [61]

    Gad-pvi: A general accelerated dynamic-weight particle-based variational inference framework

    Wang, F., Zhu, H., Zhang, C., Zhao, H., and Qian, H. Gad-pvi: A general accelerated dynamic-weight particle-based variational inference framework. Entropy, 26 0 (8): 0 679, 2024 b

    2024

  54. [62]

    Theoretical insights for diffusion guidance: A case study for gaussian mixture models

    Wu, Y., Chen, M., Li, Z., Wang, M., and Wei, Y. Theoretical insights for diffusion guidance: A case study for gaussian mixture models. arXiv preprint arXiv:2403.01639, 2024

    2024 arXiv

  55. [63]

    Lipschitz singularities in diffusion models

    Yang, Z., Feng, R., Zhang, H., Shen, Y., Zhu, K., Huang, L., Zhang, Y., Liu, Y., Zhao, D., Zhou, J., et al. Lipschitz singularities in diffusion models. In The Twelfth International Conference on Learning Representations, 2023

    2023

  56. [64]

    Functional analysis, volume 123

    Yosida, K. Functional analysis, volume 123. Springer Science & Business Media, 2012

    2012

  57. [65]

    P., and Kleijn, W

    Zhang, G., Lewis, J. P., and Kleijn, W. B. Exact diffusion inversion via bi-directional integration approximation. arXiv preprint arXiv:2307.10829, 2023

    2023 arXiv

  58. [66]

    Formulating discrete probability flow through optimal transport

    Zhang, P., Yin, H., Li, C., and Xie, X. Formulating discrete probability flow through optimal transport. Advances in Neural Information Processing Systems, 36, 2024 a

    2024

  59. [67]

    Tackling the singularities at the endpoints of time intervals in diffusion models

    Zhang, P., Yin, H., Li, C., and Xie, X. Tackling the singularities at the endpoints of time intervals in diffusion models. arXiv preprint arXiv:2403.08381, 2024 b

    2024 arXiv

  60. [68]

    Improved techniques for maximum likelihood estimation for diffusion ODE s

    Zheng, K., Lu, C., Chen, J., and Zhu, J. Improved techniques for maximum likelihood estimation for diffusion ODE s. In Krause, A., Brunskill, E., Cho, K., Engelhardt, B., Sabato, S., and Scarlett, J. (eds.), Proceedings of the 40th International Conference on Machine Learning,...

    2023

  61. [69]

    Analyzing and improving model collapse in rectified flow models

    Zhu, H., Wang, F., Ding, T., Qu, Q., and Zhu, Z. Analyzing and improving model collapse in rectified flow models. arXiv preprint arXiv:2412.08175, 2024 a

    2024 arXiv

  62. [70]

    Neural sinkhorn gradient flow

    Zhu, H., Wang, F., Zhang, C., Zhao, H., and Qian, H. Neural sinkhorn gradient flow. arXiv preprint arXiv:2401.14069, 2024 b

    2024 arXiv

  63. [71]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.