Pith. sign in

REVIEW 4 major objections 5 minor 43 references

Intrinsic-Hybrid Latent Diffusion Models for Generative Modeling on Unknown Manifolds

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ILDM claims that running diffusion in a latent chart of an unknown Riemannian manifold, switching between Riemannian and Euclidean Brownian motion according to decoder uncertainty, improves generation quality over standard and latent…

desk verdict A sensible geometry-aware latent diffusion idea that currently lacks a validated reverse sampler, controlled approximations, and sufficient empirical evidence; deserves review only if the missing appendices and experiments are supplied. read the letter →

arxiv 2608.04827 v1 pith:4M6FEMET submitted 2026-08-05 stat.ML cs.LG

Yizhu Wang , Mu Niu , Xiaochen Yang This is my paper
classification stat.MLcs.LG MSC 58J6560J60
keywords unknownmanifoldslatentdiffusionmodelsRiemannianBrownianmotionscorematchingmetricprobabilisticdecoderssparse-datagenerationhybrid
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

This paper introduces ILDM, a generative model aimed at settings with little training data where ordinary diffusion models degrade. The key move is to treat the latent space of a pretrained probabilistic decoder as a chart of an unknown Riemannian manifold, so that the geometry and uncertainty of the mapping into data space enter the diffusion dynamics. The forward process alternates between Riemannian Brownian motion in regions where the decoder is confident and ordinary Euclidean dynamics where it is not. Scores are learned from simulated hybrid trajectories through an approximate denoising score matching objective, and sampling runs a hybrid backward Langevin process. On COIL-100, MNIST, and cardiac MRI subsets, ILDM reports lower FID and LPIPS than standard diffusion and latent diffusion baselines.

What carries the argument

The machinery is the hybrid forward and backward diffusion on a latent data noise manifold. The Riemannian metric $g=J^{\top}J$, with the noncentral Wishart mean used for GP decoders, determines the local geometry, and the decoder's predictive variance $\sigma^2(x)$ determines the switching threshold $\sigma^2_{\text{thresh}}=\alpha\sigma^2_{\max}$. Riemannian Brownian motion is simulated in low-uncertainty regions, Euclidean Brownian motion in high-uncertainty regions, and the score network is trained with the ADSM loss, which replaces the intractable $p_{0t}(x_t|x_0)$ with $N(\tilde{\mu}_t, \tilde{\tau}_t^2)$ computed from simulated paths. Backward sampling uses Riemannian Langevin dynamics with natural gradient $g^{-1}s_\theta$ in confident regions and Euclidean Langevin dynamics elsewhere.

What would settle it

Run the hybrid forward process on a two-dimensional synthetic manifold with known geometry, record many conditional samples $x_t|x_0$, and compare their empirical distribution with the Gaussian used by ADSM. A growing divergence in Wasserstein distance as $t$ increases, or a strong dependence on the switching threshold $\alpha$, would show the score estimator is biased and the backward process cannot restore the true data distribution.

Watch

Extended reading notes

Core claim

The central discovery is that making the latent diffusion process aware of the data manifold's geometry and of the decoder's uncertainty yields better generative quality in sparse-data regimes than Euclidean latent diffusion or image-space diffusion. Concretely, ILDM interprets the latent space as a chart of an unknown Riemannian manifold, builds the metric from the Jacobian of a pretrained probabilistic decoder, and runs a forward hybrid SDE: Riemannian Brownian motion when the decoder's predictive variance is below a threshold, and Euclidean Brownian motion otherwise. Because the hybrid transition density is intractable, the paper approximates the conditional density by a Gaussian whose mean and variance are estimated from simulated trajectories and uses it to define an approximate denoising score matching objective. The backward process is correspondingly hybrid Riemannian-Euclidean Langevin dynamics. In experiments, the model achieves lower FID and LPIPS than DM and LDM on the Lucky Cat subset of COIL-100, cardiac MRI slices, and a 120-image MNIST subset.

Load-bearing premise

The method depends on the hybrid forward process being well approximated by a Gaussian whose mean and variance come from a handful of simulated paths; if that approximation is off, the learned score field is biased and sampling will miss the true data distribution.

Editorial extensions

If this is right

  • In data-sparse regimes, latent diffusion can be improved without extra data by exploiting decoder geometry and uncertainty rather than assuming a flat latent space.
  • Intrinsic diffusion can be applied to unknown manifolds without requiring Laplace-Beltrami operators or explicit manifold structure, since the geometry is derived from a pretrained probabilistic decoder.
  • The ADSM objective provides a way to train score networks for forward processes whose transition densities are intractable, provided simulated trajectories are available.
  • The switching threshold $\alpha$ is tunable, and the reported gains hold across $\alpha \in \{0.3, 0.4, 0.5\}$, with all settings beating standard LDM on FID.
  • Class-conditional generation is available through classifier-free guidance over the learned conditional score field.

Reading between the lines

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

  • The paper keeps the decoder fixed during diffusion training; an extension it leaves implicit is fine-tuning the decoder jointly with the score network, which could tighten the metric exactly where the backward process needs it.
  • Because the switching rule depends only on predictive variance, the same hybrid mechanism could be dropped into other latent generative models, such as flow matching, rather than only score-based diffusion.
  • A testable extension is to use the fraction of time a backward trajectory spends in high-uncertainty regions as an out-of-distribution score, since those regions are traversed only when the manifold structure is uninformative.
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

4 major / 5 minor

Summary. The paper proposes the Intrinsic Hybrid Latent Diffusion Model (ILDM), a generative framework that combines probabilistic dimensionality reduction with geometry-aware diffusion in a learned latent space. A pretrained decoder defines a Riemannian metric from the Jacobian of the latent-to-data mapping, and the forward process alternates between Riemannian Brownian motion in low-uncertainty regions and Euclidean Brownian motion in high-uncertainty regions, with the switch based on a threshold on the decoder variance. Because the transition density of this hybrid SDE is intractable, the paper introduces an approximate denoising score matching (ADSM) objective in which per-x0 transition densities are replaced by Gaussians whose moments are estimated from simulated forward trajectories. The backward process is specified as a hybrid of Euclidean and Riemannian Langevin dynamics driven by the learned scores. Experiments on COIL-100 'lucky cat', cardiac MRI, and a 120-image MNIST subset report lower FID and LPIPS scores for ILDM than for standard diffusion (DM) and latent diffusion (LDM) baselines.

Significance. If the method's theoretical basis were established, the idea of adaptively switching between metric-aware and uncertainty-aware diffusion in latent spaces would be a useful contribution to generative modeling in sparse-data settings. The paper identifies a real limitation of Euclidean LDMs and offers a concrete, implementable algorithm; the qualitative vector-field comparisons and the reported FID/LPIPS improvements are suggestive. However, the contribution is currently conditioned on two unproven approximations: the hybrid backward process is not shown to reverse the forward SDE, and the Gaussian ansatz in ADSM is uncontrolled. The paper provides no code, no repeated-run error bars, and several referenced appendices are absent from the manuscript, so reproducibility is limited. The empirical gains are modest in some comparisons and are not statistically supported.

major comments (4)
  1. [Section III-C, Eqs. (8)-(10)] The backward process is not derived as the time-reversal of the forward hybrid SDE (4)-(5). For a fixed-Riemannian forward SDE, the reverse-time drift contains both a score term and metric-dependent correction terms; Eq. (9) omits the G^{-1/2} ∂(g^{-1} G^{1/2})/∂x term that Eq. (10) appends, and no derivation connects these equations to the forward system. For the switching process, the time-reversal must additionally account for the switch-region boundary and for the possibility that the forward regime at a given latent location differs from the backward regime at the same location. Algorithm 2's update rules, referenced as Eq. (12) and Eq. (13), do not appear in the manuscript, and no stationary-density check is reported. Since the backward sampler produces the images that are scored in Tables I-III, this gap is load-bearing for the central empirical claim.
  2. [Section III-B, Eq. (6)] The ADSM objective fits the score network to the score of a per-x0 Gaussian N(μ̃_t, τ̃_t^2), with moments estimated from simulated trajectories. Even if these conditional Gaussians matched the true transition marginals, denoising score matching with a Gaussian ansatz does not yield an unbiased estimate of the marginal score ∇ log p_t when the true transition density is non-Gaussian; the approximation error is never quantified. The paper itself states that the hybrid transition density is intractable, making this an uncontrolled approximation. The claimed agreement between ADSM and SSM is supported only by qualitative vector-field plots, not by a numerical score-error or distributional comparison. Because the learned score drives the backward dynamics, this approximation is load-bearing for the validity of the sampler.
  3. [Tables I-III and Section IV] All quantitative comparisons are single-run evaluations without error bars, confidence intervals, or multiple random seeds, and FID/LPIPS are computed on only 200 generated samples. Several reported gains are modest, such as LPIPS 0.5049 vs. 0.5442 in Table I and LPIPS 0.6816 vs. 0.6891 in Table III. The abstract's claim that ILDM 'significantly improves' generation quality is therefore not supported by the evidence as presented. The switching threshold α is described as tunable, but its chosen values per dataset are not reported, and the ablation in Appendix B is not included in the manuscript.
  4. [Appendices and Algorithm 2] Several load-bearing implementation details are deferred to appendices that are absent from the submitted text: the discretization of Eq. (10) (Appendix A), the α ablation and SSM comparison (Appendix B), GP/VAE metric construction (Appendix H), baseline SDE specifications (Appendix I), and class-conditioning algorithms (Appendix C). Algorithm 2 additionally refers to Eq. (12) and Eq. (13), which are not defined anywhere in the manuscript. As submitted, the method is not fully reproducible, and the Euler-Maruyama discretizations cannot be checked against the continuous-time SDEs.
minor comments (5)
  1. [Section IV-C] The text refers to 'the Riemannian LDM' when discussing the proposed method; this should be 'ILDM' for consistency with the tables and abstract.
  2. [Equation (6)] The notation in Eq. (6) is confusing: sθ(xt,t)·τ̃_t is a vector scaled by τ̃_t, while the target uses (xt-μ̃_t)/τ̃_t; please clarify the intended weighting and ensure the objective is dimensionally consistent.
  3. [Section III-B] The statement that the ADSM score field 'closely aligns' with the SSM score field is not supported by any quantitative comparison; a numerical score-error metric or an FID/LPIPS table comparing ADSM and SSM would be needed to substantiate this claim.
  4. [References] References [10] and [11] both appear as 'Riemannian diffusion models' in the bibliography; please disambiguate them by author groups or full titles to avoid citation confusion.
  5. [Algorithm 1] Algorithm 1 does not specify the step size Δt_f, number of starting points n_f, or number of trajectories n_BM used in the experiments; these are needed for reproducibility and for assessing whether the Euler-Maruyama discretization is in a stable regime.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central FID/LPIPS claims are measured against external benchmarks, and the ADSM objective is an explicitly stated approximation rather than a reduction of a predicted quantity to a fitted input.

full rationale

The paper's central empirical claim is that ILDM achieves lower FID and LPIPS than standard DM and LDM on three datasets; FID and LPIPS are external perceptual metrics computed on generated samples, and no fitted parameter of the model is used in these metrics. The score network is trained with the ADSM objective in Eq. (6), where the paper explicitly states 'We approximate the intractable transition density p_{0t}(x_t|x_0) by a Gaussian N(mu~_t, tau~_t^2)' using sample moments from simulated hybrid forward trajectories. This is a stated modeling approximation, not a hidden reuse of the evaluation target. The Riemannian Brownian motion SDE (Eqs. 4-5) and the hybrid switching rule are defined from the decoder-derived metric g (Eqs. 2-3), and the backward SDEs (Eqs. 8-10) are presented as heuristic constructions; whether the backward process is the true time-reversal of the forward SDE is a correctness and validation concern, not a circularity. No load-bearing step in the derivation reduces by construction to its own inputs, and there are no self-citations carrying the argument; the cited Riemannian-geometry and score-matching results are standard external references. The only tuning-related concern is that the switching threshold alpha is selected via cross-validation, which is ordinary hyperparameter selection and does not make the reported FID comparison circular.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The 'data noise manifold' is a defining concept, not an entity with independent evidence; the central free parameter is the switching threshold α, and the ADSM/Gaussian approximation and backward dynamics are the main unproven assumptions.

free parameters (1)
  • Switching threshold factor α = tested in {0.3, 0.4, 0.5}; final per-dataset value not stated
    α sets σ²_thresh = α σ²_max and controls the hybrid switch between Riemannian and Euclidean dynamics. The paper tests three values and says it can be cross-validated, but the tables do not report which α produced the results, so the evaluation may include tuning on the test metric.
assumptions (4)
  • domain assumption The decoder φ defines a smooth immersion, so g = JᵀJ (or its noncentral Wishart mean) is a valid Riemannian metric on the latent space.
    Section II, Eqs (2)-(3): the entire construction of the Riemannian BM (4)-(5) depends on the metric being smooth and positive-definite; a rank-deficient or rough Jacobian would break the SDEs.
  • ad hoc to paper For each x0, the transition density of the hybrid forward process is approximately Gaussian with moments estimated from simulated trajectories.
    Section III-B, ADSM objective (6): the authors replace the intractable p_{0t}(x_t|x_0) with N(μ̃_t, τ̃_t²), whose mean/variance come from a finite number of simulated paths. No error bound or validation is given.
  • ad hoc to paper The hybrid backward Langevin dynamics converge to the data distribution when driven by the learned scores.
    Section III-C, Eqs (8)-(10): the reverse sampler is a heuristic combination of Euclidean and Riemannian Langevin dynamics. The paper does not prove that this finite-time annealing procedure is the reverse of the forward hybrid SDE or that it samples p0.
  • domain assumption The decoder's predictive variance σ²(x) is a faithful indicator of metric reliability, so switching at α σ²_max preserves useful geometry.
    Section III-A: the switching rule assumes that variance below the threshold means the metric is trustworthy and above it means the metric is not; the paper supplies no empirical or theoretical check of this correlation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Intrinsic-Hybrid Latent Diffusion Models for Generative Modeling on Unknown Manifolds." pith.science (2026). https://pith.science/paper/4M6FEMET

@misc{pith2026260804827,
  author       = {Pith},
  title        = {Pith review of: Intrinsic-Hybrid Latent Diffusion Models for Generative Modeling on Unknown Manifolds},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4M6FEMET}},
  note         = {Machine review of arXiv:2608.04827}
}
read the original abstract

We introduce the Intrinsic Hybrid Latent Diffusion Model (ILDM), a generative framework that integrates probabilistic dimensionality reduction with geometry-aware diffusion on unknown manifolds. While diffusion models (DMs) have achieved state-of-the-art results in high-dimensional data synthesis, they rely on large training datasets and ignore intrinsic geometric structure. Latent diffusion models (LDMs) address the high dimensionality by learning a latent space, but they typically impose a Euclidean structure, failing to capture the underlying manifold geometry, especially problematic in data-sparse regimes. ILDM addresses these limitations by interpreting the latent space as a chart of an unknown Riemannian manifold, with geometry and uncertainty quantified through a probabilistic decoder. The forward process is a hybrid diffusion that switches between Riemannian and Euclidean dynamics based on local uncertainty, where the Riemannian component is governed by a probabilistic metric tensor derived from the decoder. To learn the generative dynamics, we introduce an approximate denoising score matching method tailored to the hybrid diffusion setting, enabling a backward process defined by hybrid Langevin dynamics. Experiments on COIL-100, MNIST, and cardiac MRI datasets demonstrate that ILDM significantly improves generation quality, achieving lower FID and LPIPS scores compared to standard diffusion and latent diffusion models.

Figures

Figures reproduced from arXiv: 2608.04827 by the authors.

Figure 1
Figure 1. Illustration of manifold parameterization for the COIL image dataset. The mapping ϕ : X → S projects latent points (blue triangles in the left sub-figure) to their corresponding images (six examples shown in the right sub-figure). Background shading in the latent space represents the uncertainty of ϕ, quantified as the variance of the mapping, which increases with distance from observed data points. tion from simula… view at source ↗
Figure 2
Figure 2. Score-based generative model on the data noise manifold. The blue triangles in each sub-figure represent the data points in the latent space. The left sub-figure shows three Brownian motion (BM) paths in the latent space. Starting from the black circle, the red and green paths follow Riemannian BM, while the yellow path initially follows Riemannian BM but switches to Euclidean BM when it enters the high uncertainty … view at source ↗
Figure 3
Figure 3. Comparison of vector field estimates from ILDM (a) and LDM (b) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of conditional vector fields estimated by ILDM (a) and LDM (b) under the ‘front view’ condition. The blue triangles on the right half of the latent space represent samples from the true conditional distribution corresponding to the front view of the object. …
Figure 5
Figure 5. Figure 5: Comparison of COIL-100 Lucky Cat image reconstructions using different methods. (a) Original input images. (b) Images generated by ILDM with approximate DSM. (c) Images generated by the LDM method. (d) Images generated by the DM method. perceptual image patch similarit…
Figure 6
Figure 6. Figure 6: Comparison of cardiac MRI image reconstructions using different generative models. (a) Original input images. (b) Images generated by the proposed ILDM method. (c) Images generated by LDM. (d) Images generated by DM. TABLE II: Performance comparison on the left ventric…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 36 canonical work pages

  1. [1]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems, vol. 33, pp. 6840–6851, 2020

    2020

  2. [2]

    Generative modeling by esti- mating gradients of the data distribution,

    Y . Song and S. Ermon, “Generative modeling by esti- mating gradients of the data distribution,” Advances in neural information processing systems, vol. 32, 2019

    work page 2019

  3. [3]

    Elucidating the design space of diffusion-based generative models,

    T. Karras, M. Aittala, and S. Laine, “Elucidating the design space of diffusion-based generative models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023

    work page 2023

  4. [4]

    Text-to-video generation,

    U. Singer et al., “Text-to-video generation,” in NeurIPS, 2023

    work page 2023

  5. [5]

    A connection between score matching and denoising autoencoders,

    P. Vincent, “A connection between score matching and denoising autoencoders,” Neural computation, vol. 23, no. 7, pp. 1661–1674, 2011

    work page 2011

  6. [6]

    Deep unsupervised learning using nonequi- librium thermodynamics,

    J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequi- librium thermodynamics,” in International conference on machine learning. pmlr, 2015, pp. 2256–2265

    work page 2015

  7. [7]

    Score-based generative mod- eling through stochastic differential equations,

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

    arXiv 2011

  8. [8]

    Test- ing the manifold hypothesis,

    C. Fefferman, S. Mitter, and H. Narayanan, “Test- ing the manifold hypothesis,” Journal of the American Mathematical Society, vol. 29, no. 4, pp. 983–1049, 2016

    work page 2016

Show all 43 references
  1. [9]

    Extrinsic gaussian processes for regression and classification on manifolds,

    L. Lin, N. Mu, P. Cheung, and D. Dunson, “Extrinsic gaussian processes for regression and classification on manifolds,” Bayesian Analysis, vol. 14, no. 3, pp. 887– 906, 2019

    2019

  2. [10]

    Riemannian diffusion models,

    C.-W. Huang, M. Aghajohari, J. Bose, P. Panangaden, and A. C. Courville, “Riemannian diffusion models,” Advances in Neural Information Processing Systems, vol. 35, pp. 2750–2761, 2022

    2022

  3. [11]

    Riemannian diffusion models,

    V . De Bortoli, J. Xu, J. Thornton et al., “Riemannian diffusion models,” in NeurIPS, 2022

    2022

  4. [12]

    Generative modeling on manifolds through mixture of riemannian diffusion processes,

    M. Jo, J. Ha et al., “Generative modeling on manifolds through mixture of riemannian diffusion processes,” in AISTATS, 2024

    2024

  5. [13]

    High-resolution image synthesis with latent diffusion models,

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High-resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

    2022

  6. [14]

    An introduction to variational autoencoders,

    D. P. Kingma, M. Welling et al., “An introduction to variational autoencoders,” Foundations and Trends® in Machine Learning, vol. 12, no. 4, pp. 307–392, 2019

    2019

  7. [15]

    Probabilistic non-linear principal com- ponent analysis with gaussian process latent variable models,

    N. Lawrence, “Probabilistic non-linear principal com- ponent analysis with gaussian process latent variable models,” Journal of machine learning research, vol. 6, no. Nov, pp. 1783–1816, 2005

    2005

  8. [16]

    Learning for larger datasets with the gaussian process latent variable model,

    N. D. Lawrence, “Learning for larger datasets with the gaussian process latent variable model,” in Artificial intelligence and statistics, 2007, pp. 243–250

    2007

  9. [17]

    Revis- 10 iting active sets for gaussian process decoders,

    P. Moreno-Mu ˜noz, C. Feldager, and S. Hauberg, “Revis- 10 iting active sets for gaussian process decoders,”Advances in Neural Information Processing Systems, vol. 35, pp. 6603–6614, 2022

    2022

  10. [18]

    Columbia object image library (coil-100),

    S. A. Nene, S. K. Nayar, H. Murase et al., “Columbia object image library (coil-100),” 1996

    1996

  11. [19]

    Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?

    O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. G. Ballester et al., “Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?” IEEE transactions on medi...

    2018

  12. [20]

    Multi- centre, multi-vendor and multi-disease cardiac image segmentation challenge,

    V . M. Campello, J. F. R. Palomares, A. Guala, M. Marakas, M. Friedrich, and K. Lekadir, “Multi- centre, multi-vendor and multi-disease cardiac image segmentation challenge,” 2020. [Online]. Available: https://zenodo.org/record/3886268

    2020

  13. [21]

    Gradient-based learning applied to document recogni- tion,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recogni- tion,” Proceedings of the IEEE, vol. 86, no. 11, pp. 2278– 2324, 1998

    1998

  14. [22]

    Metrics for probabilistic geometries,

    A. Tosi, S. Hauberg, A. Vellido, and N. D. Lawrence, “Metrics for probabilistic geometries,” arXiv preprint arXiv:1411.7432, 2014

    2014 arXiv

  15. [23]

    Fast and robust shortest paths on manifolds learned from data,

    G. Arvanitidis, S. Hauberg, P. Hennig, and M. Schober, “Fast and robust shortest paths on manifolds learned from data,” in The 22nd International Conference on Artificial Intelligence and Statistics. PMLR, 2019, pp. 1506–1515

    2019

  16. [24]

    Laplacian eigenmaps for dimensionality reduction and data representation,

    M. Belkin and P. Niyogi, “Laplacian eigenmaps for dimensionality reduction and data representation,”Neural computation, vol. 15, no. 6, pp. 1373–1396, 2003

    2003

  17. [25]

    Nonlinear dimensionality reduction by locally linear embedding,

    S. T. Roweis and L. K. Saul, “Nonlinear dimensionality reduction by locally linear embedding,” science, vol. 290, no. 5500, pp. 2323–2326, 2000

    2000

  18. [26]

    J. M. Lee, Introduction to smooth manifolds, 2nd ed., ser. Graduate Texts in Mathematics. Springer, New York, 2013, vol. 218

    2013

  19. [27]

    C. E. Rasmussen and C. K. Williams, Gaussian processes for machine learning. MIT press Cambridge, MA, 2006, vol. 2, no. 3

    2006

  20. [28]

    The non-central wishart distribution and certain problems of multivariate statistics,

    T. W. Anderson, “The non-central wishart distribution and certain problems of multivariate statistics,” The Annals of Mathematical Statistics, pp. 409–431, 1946

    1946

  21. [29]

    Brownian motion and Riemannian geometry,

    P. Hsu, “Brownian motion and Riemannian geometry,” Contemporary Mathematics, vol. 73, pp. 95–104, 1988

    1988

  22. [30]

    A brief introduction to Brownian motion on a Riemannian manifold,

    E. P. Hsu, “A brief introduction to Brownian motion on a Riemannian manifold,” Lecture Notes, 2008

    2008

  23. [31]

    Higher-order implicit strong numerical schemes for stochastic differential equa- tions,

    P. E. Kloeden and E. Platen, “Higher-order implicit strong numerical schemes for stochastic differential equa- tions,” Journal of Statistical Physics, vol. 66, no. 1, pp. 283–314, 1992

    1992

  24. [32]

    Lamberton and B

    D. Lamberton and B. Lapeyre, Introduction to Stochastic Calculus Applied to Finance. CRC press, 2007

    2007

  25. [33]

    Estimation of non- normalized statistical models by score matching

    A. Hyv ¨arinen and P. Dayan, “Estimation of non- normalized statistical models by score matching.” Journal of Machine Learning Research, vol. 6, no. 4, 2005

    2005

  26. [34]

    U- net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U- net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18. ...

    2015

  27. [35]

    Riemann manifold langevin and hamiltonian monte carlo methods,

    M. Girolami and B. Calderhead, “Riemann manifold langevin and hamiltonian monte carlo methods,” Journal of the Royal Statistical Society Series B: Statistical Methodology, vol. 73, no. 2, pp. 123–214, 2011

    2011

  28. [36]

    Stochastic gradient hamiltonian monte carlo,

    T. Chen, E. Fox, and C. Guestrin, “Stochastic gradient hamiltonian monte carlo,” in International conference on machine learning. PMLR, 2014, pp. 1683–1691

    2014

  29. [37]

    Amari and H

    S.-i. Amari and H. Nagaoka, Methods of information geometry. American Mathematical Soc., 2000, vol. 191

    2000

  30. [38]

    Classifier-free diffusion guid- ance,

    J. Ho and T. Salimans, “Classifier-free diffusion guid- ance,” arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  31. [39]

    Diffusion models beat gans on image synthesis,

    P. Dhariwal and A. Nichol, “Diffusion models beat gans on image synthesis,” Advances in neural information processing systems, vol. 34, pp. 8780–8794, 2021

    2021

  32. [40]

    Progressive distillation for fast sampling of diffusion models,

    T. Salimans and J. Ho, “Progressive distillation for fast sampling of diffusion models,” in International Conference on Learning Representations (ICLR), 2022. [Online]. Available: https://openreview.net/forum?id= TIdIXIpzhoI

    2022

  33. [41]

    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,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  34. [42]

    The unreasonable effectiveness of deep fea- tures as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang, “The unreasonable effectiveness of deep fea- tures as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 586–595

    2018

  35. [43]

    The intrinsic dimension of images and its impact on learning,

    P. Pope, C. Zhu, A. Abdelkader, M. Goldblum, and T. Goldstein, “The intrinsic dimension of images and its impact on learning,” in Proceedings of the 9th International Conference on Learning Representations (ICLR), 2021. [Online]. Available: https://openreview. net/forum?id=XJk19XzGq2J

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.