acceptodds
Under review as a conference paper at ICLR 2027

Diffusion Processes on Implicit Manifolds

Abstract

Simulating diffusion processes on a manifold typically requires access to tangent spaces and projection operators, which are usually unavailable when the manifold is known only through samples. We introduce Implicit Manifold-valued Diffusions (IMDs): a data-driven framework for constructing continuous stochastic dynamics on such implicit manifolds. Rather than reconstructing the geometry explicitly, IMDs estimate a manifold diffusion operator from the point cloud and use it to define the drift and diffusion coefficients of an ambient-valued SDE. The point cloud therefore serves only to estimate the diffusion operator. The induced process evolves continuously in the ambient space and becomes manifold-valued in the large-sample limit. We prove that strong-resolvent convergence of the estimated operator is sufficient for weak convergence of the induced stochastic process on path space. As a concrete realization, we construct a family of graph-based estimators whose induced IMDs range from density-biased Langevin dynamics to manifold Brownian motion. Experiments validate the predicted path space convergence and demonstrate geodesic recovery, density-debiased manifold exploration, and stochastic interpolation on image data.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
diffusion processes, manifold hypothesis, stochastic differential equations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_migQOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces Implicit Manifold-valued Diffusions (IMDs), a framework for approximating diffusion processes on manifolds known only through point samples. It constructs ambient-space SDEs whose drift and covariance are derived from an estimated diffusion generator and its carré du champ. The main theoretical claim is that strong-resolvent convergence of the estimated operators ensures weak convergence of the induced processes on path space. A diffusion-maps-style graph construction yields a family ranging from density-biased Langevin dynamics to manifold Brownian motion. Numerical validation includes geodesic estimation on an hourglass surface, path-distribution and geometric errors on the Stanford Bunny, density debiasing on a torus, and guided COIL-20 image interpolation using learned viewpoint augmentation.

Critical strengths
  1. A coherent operator-based construction. Deriving drift and covariance from an estimated generator and its carré du champ provides a clear connection between graph-based manifold learning and continuous SDE simulation. The density-correction parameter also gives an interpretable family of dynamics. This construction is potentially useful for stochastic simulation when the manifold is available only through samples.

  2. Evaluation aligned with the dynamical objective. The experiments assess path MMD, quadratic variation, and off-manifold error across sample sizes. These complementary diagnostics examine temporal behavior and geometric fidelity, providing meaningful empirical evidence relevant to the paper’s central goal of approximating entire diffusion processes.

Critical weaknesses
  1. The central path-space convergence guarantee is not established by the provided proof. The Taylor comparison in Proposition 3 does not justify transferring strong-resolvent convergence of the graph operator to the interpolated continuous SDE. Appendix C nevertheless uses the graph semigroup to represent IMD expectations in Eq. (69). It also infers convergence from a fixed starting point from L2L^2 semigroup convergence and asserts the uniform increment bound in Eq. (71) without deriving it. These gaps affect Theorem 1, which underpins the paper’s principal theoretical contribution.

  2. The real-data experiment does not isolate the benefit of IMDs. The COIL-20 pipeline combines learned viewpoint augmentation, strong endpoint guidance, tangent-plane retraction, and noise tapering (Appendix F.5), but compares only against linear interpolation. Without baselines using the same augmented data and ablations of the added components, the reported improvement cannot be attributed specifically to the proposed diffusion construction. This substantially limits the evidence for its practical advantage on real data.

Questions for authors
  1. Clarification of the graph-to-SDE convergence argument. How does the Taylor estimate in Proposition 3 imply semigroup convergence for the continuous IMD after off-sample interpolation? In particular, how is using the graph semigroup etL^Ne^{t\widehat L_N} for IMD expectations in Eq. (69) justified, given that the continuous process has generator GNG_N?

  2. Clarification of initial conditions and tightness. What assumptions justify convergence from a fixed starting point from the L2L^2 semigroup limit, and how is the NN-uniform increment bound in Eq. (71) derived? If additional assumptions are required, are they satisfied by the graph construction and interpolation used in the experiments?

  3. Clarification of the COIL-20 evidence. Are matched baselines or ablations available that separate the contribution of IMD dynamics from learned viewpoint augmentation and the additional bridge corrections? Results using the same augmented data and comparable endpoint guidance would materially affect my assessment of the method’s practical contribution.

2 comments

We thank the reviewer for providing insightful feedback, which, upon revision, has greatly improved the clarity and tightness of the manuscript’s theoretical statements and experiments.

How does the Taylor estimate in Proposition 3 imply semigroup convergence for the continuous IMD after off-sample interpolation?

We thank the reviewer for this comment. We believe the confusion arises from conflating Theorem 1 and the graph-based estimator in Section 6. Crucially, IMDs are defined through operators, making Prop. 2 an independent, rigorous asymptotic result. Let us now clarify that the purpose of the Taylor expansion was to motivate the finite-NN continuous IMD construction. It was not intended to establish semigroup convergence. The latter follows from Prop. 2, which relies on generalized strong-resolvent convergence. We have clarified the intent of Prop. 3, and added a short discussion relating to existing literature, e.g. [2,3] in Appx. B.5.

In particular, how is using the graph semigroup for IMD expectations in Eq. (69) justified, given that the continuous process has generator GNG_N?

This question relates to the previous, and arose from a typo in the proof of Theorem 1. Introducing GNG_N was confusing, and led to conflating it with Proposition 3. We reiterate that Theorem 1 is based on a general estimated operator L^N\hat{\mathcal{L}}_N, not on any particular graph Laplacian or semigroup. We have now fixed this typo.

Are matched baselines or ablations available that separate the contribution of IMD dynamics from learned viewpoint augmentation and the additional bridge corrections? Results using the same augmented data and comparable endpoint guidance would materially affect my assessment of the method’s practical contribution.

We agree with the reviewer that the viewpoint augmentation did not adequately capture the contribution of IMDs. Prompted by this critique, we have now updated the image dataset experiment of Sec. 7.4 to better highlight the benefit of IMDs by generating a controllable synthetic dataset, thereby removing the augmentation, providing both a ground truth and a baseline (Fig. 6).

In particular, we extended the dataset to 2-dimensional rotations, effectively augmenting the implicit manifold from a circle to a torus. We conducted a comparison of pixel-space IMDs to image-rendered toroidal Brownian motion (ground truth) and a graph-based random walk (baseline). See Figs. 6, 9 and 10 of the updated manuscript. We observe that IMDs reproduce the reference diffusion's path MMD and QV with accuracy comparable to the graph random walk, while generating continuous trajectories in pixel space. This demonstrates that IMDs perform Brownian motion in the high-dimensional space of images of rotated 3D objects (Table 5). This better captures the contribution of IMDs as a method for stochastic dynamics on data manifolds without geometric priors. We also performed kernel-type ablations in Fig. 10. We refer the reviewer to Section 7.4, and Appx. F.5.1 of the updated manuscript for further details.

What assumptions justify convergence from a fixed starting point from the semigroup limit, and how is the N-uniform increment bound in Eq. (71) derived?

Thank you for this question. We believe this arises from a typo in the proof of Theorem 1. Specifically, Theorem 1 proves convergence for a sampled starting point and not a fixed starting point. We have corrected this issue.

The increment bound follows from standard results in stochastic analysis, particularly Hölder's inequality and the Burkholder-Davis-Gundy inequality [1]. These allow us to control the increments of the stochastic process using the uniform moment bounds on its drift and diffusion coefficients, which are now explicitly stated before the proof of Theorem 1.

If additional assumptions are required, are they satisfied by the graph construction and interpolation used in the experiments?

All assumptions required to prove the asymptotic graph-operator limit (Prop. 2) are in Assumption 1 (Appx. B). We emphasize that transferring this convergence to the particular interpolated continuous SDEs used in our experiments requires additional conditions that are typical in established graph-Laplacian consistency and stochastic process approximation results [2,3]. We discuss these in Appx. B.5.

Furthermore, the experiments complement our theory by demonstrating practical effectiveness of our method. We believe this combination of rigorous asymptotic analysis and empirical validation provides a solid foundation for IMDs, with further guarantees for finite-sample dynamics, making for a natural direction for future work.

[1] Revuz and Yor, Continuous martingales and Brownian motion

[2] Guérin, Nguyen and Tran, Strong Uniform Convergence of Laplacians of Random Geometric and Directed kNN Graphs on Compact Manifolds

[3] Ethier and Kurtz, Markov Processes: Characterization and Convergence

AuthorsOpenReview ↗

Followup on our rebuttal

Dear Reviewer migQ, thank you again for your time and constructive feedback.

We have addressed your concerns by clarifying and fixing the typos in the proofs, as well as extending our image experiment to include both a ground truth and more suitable baseline, better isolating the capabilities of IMDs. Since the discussion period will soon be coming to an end, we would appreciate it if you could consider our rebuttal and improved manuscript and take these changes into account in your assessment.

We would be happy to respond to any further questions or concerns you may have.

Reviewer_6737OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on how to simulate continuous stochastic processes on manifolds formed by sampled data points, where exact geometric properties such as exp maps and geodesics are unknown. The main difficulty stated by the paper is that existing manifold SDE methods require explicit geometric to be known, while graph random walks only move between sampled points instead of producing continuous trajectories. To address this, the paper proposes Implicit Manifold-valued Diffusions (IMDs), which estimate a diffusion operator by forming a graph over the sampled data and use it to construct the drift and diffusion coefficients of an ambient space SDE. They theoretically show that strong resolvent convergence of the estimated operator is sufficient for weak convergence of the resulting stochastic processes on path space. They also prove that their graph estimator converges to a family of manifold diffusion operators. They evaluate IMDs on on the following tasks: hourglass surface for predicting geodesic distances, the Stanford Bunny, a torus for exploration with sample density bias, and COIL-20 images for stochastic interpolation. The experiments show that IMDs can predict geodesic distances, better match the reference Brownian dynamics as the number of samples grows, and achieve better manifold coverage when removing density bias. On COIL-20, IMDs also outperform linear interpolation.

Critical strengths
  1. The main strength of the paper is its framework for constructing continuous stochastic processes on implicit manifolds using diffusion operators estimated from data. This allows the method to generate continuous trajectories beyond the observed samples, instead of being limited to random walks on a graph. The authors also provide theoretical results connecting operator convergence to convergence of entire stochastic trajectories, which helps justify the proposed construction.
  2. The experimental results generally supports the method, such as the Stanford Bunny results showing improved performance as the number of samples increases. These results provide reasonable validation that the framework works, although they do not establish its advantages over existing methods.
Critical weaknesses
  1. The main weakness of the paper is that it does not provide sufficient comparisons with prior methods that address similar problems. Many recent works have attempted to learn transport dynamics and trajectory dynamics in implicit, unknown data manifolds, many of which also perform experiments on tasks similar to those in this paper. Just some examples include works that tackle un-biased exploration and generation on unknown data manifolds [b, c], geodesic prediction, latent interpolation (especially on images), and trajectory inference [a, b, e]. RFM [d] also evaluates geometric generative modeling on the Stanford Bunny, which IMD uses in its experiments. While some of these do not focus the exact same technicalities as IMD, the eventual experimental application are very closely related. Despite these closely related works, the paper mainly compares against simple baselines such as ambient Brownian motion and linear interpolation, reference processes that use known manifold geometry, or different variants of IMD itself. While these experiments demonstrate that IMD can perform the proposed tasks, they do not establish how well it performs compared to existing geometric methods. This is particularly important because these prior works share the same broader goal of sampling, generation, or interpolation on implicit manifolds, even if they use different mathematical formulations. Without these comparisons, it is difficult to assess the concrete contribution of the proposed framework over existing methods. While comparing against every prior method may be difficult, a careful discussion of IMD's concrete advantages and selective experimental comparisons would still greatly strengthen the paper.

[a] Kacper Kapuśniak, et al. Metric Flow Matching for Smooth Interpolations on the Data Manifold

[b] Sun et al. Geometry-Aware Generative Autoencoders for Warped Riemannian Metric Learning and Generative Modeling on Data Manifolds

[c] De Santi et al. Provable Maximum Entropy Manifold Exploration via Diffusion Models

[d] Chen and Lipman. Flow Matching on General Geometries

[e] Petrovic et al. Curly Flow Matching for Learning Non-gradient Field Dynamics

  1. It seems to me that IMD could be sensitive to the quality of graph construction. The paper doesn't perform any ablation on the hyperparameters on how the graph is constructed (e.g. neighborhood size, the number of neighbors used interpolate drift and diffusion coefficients, the kernel bandwidth, kernel types, etc). This makes it hard to assess the robustness of the method.
Questions for authors
  1. The authors mention that they expect Gaussian kernels to satisfy the same theoretical convergence guarantees as hard cutoff kernels, but do not provide much justification. Could the authors explain intuitively why they expect this to hold? Also, since the COIL-20 experiment uses a Gaussian kernel instead of the hard cutoff kernel used in the theoretical analysis, how do the performance of these two kernel compare with each other on the task?
3 comments

We thank the reviewer for their feedback on our theoretical work and experimental validation. The reviewer’s concern was predominantly regarding differentiating IMDs from existing methods, which we address below, along with other comments.

The main weakness of the paper is that it does not provide sufficient comparisons with prior methods that address similar problems.

We thank the reviewer for this critique. We believe this arises from the lack of clarity on the specific advantages of IMDs over other methods, which we address below.

Let us reiterate that the main contribution of IMDs is a data-driven construction of continuous-time, manifold-valued diffusions on implicit manifolds, with pathwise convergence guarantees (Thm 1, Prop. 2). To our knowledge, no purely data-driven method for constructing stochastic trajectories on manifolds has pathwise convergence guarantees. Unlike existing task-specific methods, IMDs construct continuous stochastic dynamics directly from sampled manifold geometry, with sufficient conditions for weak convergence on path space. Existing methods mostly focus on distributional guarantees at fixed times, and typically require analytic access to the manifold.

The closest work we are aware of [1] uses a pre-trained score model to recover manifold retraction and projection operations. IMDs differ in that they leverage geometric stochastic structure through L,Γ\mathcal L, \Gamma, enabling the construction of canonical manifold-valued stochastic processes directly from data.

Necessarily, pathwise convergence implies distributional convergence. In Fig. 4b, we show that IMDs outperform PF-ODEs (a baseline method with distributional but not path-space convergence guarantees) by matching the ground truth process in MMD and QV.

While IMDs differ from all other methods in their path space guarantees, we additionally compare IMDs to the methods cited by the reviewer from the viewpoint of shared applications, where distributional guarantees are sufficient. Specifically,

  • [b,c] target uniform sampling on M\mathcal M. IMDs achieve this by setting α=1\alpha=1 (Tab. 2). We compare against [b] and [c] in Table 4, finding a 200-fold shorter runtime for IMDs than [b] with very similar performance. Meanwhile, [c] failed to outperform IMDs on all metrics. See lines 423 and 458 of the updated manuscript.
  • [d] requires the manifold explicitly, which is not provided in our setting.
  • [a] constructs deterministic interpolants between distributions, therefore addressing a different problem than IMDs.
  • [e,4] construct stochastic processes with a reference diffusion. These methods could consider an IMD as the reference when on an implicit manifold. We see this as the natural route for task-level methods to build on IMDs.

We clarify the nuances in an updated Related Works section, and have also updated our experiments to compare with selected current methods.

Could the authors explain intuitively why they expect [the use of Gaussian kernels] to hold? The paper doesn't perform any ablation on the hyperparameters on how the graph is constructed.

Both cut-off and Gaussian kernels capture local interactions, but the cut-off kernel truncates, while Gaussians smoothly exponentially dampen distant observations. Refs. [2,3] support the idea that both are expected to recover the same limiting differential operator. Yet, Gaussian kernels have unbounded support, which presents additional technical challenges. We conjecture that extending our proofs to Gaussian kernels would be possible using techniques in [2,3], but this is beyond the scope of this work. We have added a discussion to Appx. B.4, l. 1173.

To experimentally address the similarity between both kernels, we conducted ablations over kernel bandwidth and kk-nn neighbours on the Stanford Bunny experiment. Both kernels performed statistically indistinguishably on the path MMD and QV metrics at default settings, and displayed comparable sensitivity to kernel bandwidth, while remaining largely unaffected by varying kk-neighbourhood. We refer to Appx. F.3 and F.5.1 of the updated manuscript for the full ablation results.

How do the performance of these two kernel compare with each other on the COIL-20 task?

We used the Gaussian kernel for COIL-20 because a local cut-off kernel left the graph disconnected due to low sample count. An updated experiment in Section 7.4 now includes the same kernel ablations, supporting our above intuition in higher-dimensional pixel spaces.

[1] Kharitenko et al., Landing with the score: Riemannian optimization through denoising

[2] García Trillos et al., Error estimates for spectral convergence of the graph Laplacian on random geometric graphs toward the Laplace-Beltrami operator

[3] Cheng and Wu. Eigen-convergence of Gaussian kernelized graph Laplacian by manifold heat interpolation

[4] De Bortoli et al., Riemannian score-based generative modelling

I thank the for their reply. However, I'm not sure the justification on the Gaussian kernel gives enough theoretical evidence that the property should hold. For a theory heavy paper, not being able to sufficiently demonstrate theoretical claims made by the paper and relying on empirical justification is in my opinion rather a substantial weakness. Furthermore, while I understand that the technical details of the method itself is different from some of the papers listed, ultimately the application of the method demonstrated by the paper are similar to these prior works. It is ultimately on the authors to demonstrate the value of performing manifold-valued diffusion on implicit manifolds as compared to say existing flow matching or schrodinger bridge method on implicit manifolds (e,g, geodesic distance estimation, path MMD, etc). I appreciate the comparison the authors added against GAGA and S-MEME (although I think claiming similar performance to GAGA is overstating it, as IDM is worse than GAGA by roughly 25-50% in 3 of the 4 tasks and far greater than the error bars), which partly addresses this problem. As a result, given the rebuttal timeline, I encourage the authors to focus on the theoretical justification if the authors are restricted on time.

We thank the reviewer for their reply.

We would like to emphasize again that our theory and the main empirical results are self-contained and are solely based on the cutoff kernel. Motivated by their frequent use in manifold learning practice [1], the ablations including Gaussian kernels (Figs. 7,8,10) are intended to support our reply to the reviewer’s request for an intuitive explanation (further discussed in Appx. B.4). Our intention was not to mathematically overclaim.

IMDs are introduced as a novel framework for extending existing methods to the setting of implicit data manifold (e.g. Schrödinger bridges), where such constructions remain largely intractable due to no assumed geometric priors. We have evaluated IMDs on distinct tasks (e.g. Secs. 7.1 & 7.3) to illustrate this generality, while recognizing that task-specific methods already exist.

Therefore, we acknowledge that GAGA achieves lower distributional and geometric errors on the manifold exploration task, which was included to highlight the computational advantage of IMDs showing (in an updated full experiment) approx. 900×900 \times lower CPU time than GAGA (despite GAGA using 24 CPU threads per seed compared to one for IMDs). We have updated Table 4 with the full results.

In light of these clarifications and the revisions made to address your concerns, we would be grateful if you would reconsider your assessment and score.

[1] Coifman and Lafon, Diffusion Maps

Reviewer_zy8WOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposed an indirect framework of simulating the diffusion on point-cloud sampled manifold. Instead of reconstructing manifold directly, this approach estimates diffusion operator only which effectively defines the the drift and diffusion coefficients of an ambient-valued SDE. This paper proved the estimator convergence and its sufficiency for weak convergence of the induced stochastic process on path space. A concrete realization is validated effectiveness by a numerical experiment. This paper studies on an interesting topic on diffusion process on implicit manifolds, and presents sound main results. This paper provides technically theoretical analysis, however, it still has some minor issues on experimental validation seen weakness section. That's the reason I give weak accept.

Critical strengths

1.This paper is well presented, format, and organized. Explicit visualization provides intuitively explanation for readers clearly. 2.This paper conducts a thorough investigation of relevant research and achieves a breakthrough in previous theoretical work. 3.The theoretical structure of this paper is reasonable and logically compact. The numerical experiments are progressive, verifying both the effectiveness of the theory and the practicality of the algorithm implementation.

Critical weaknesses

1.This paper involves highly technical theoretical research, resulting in a relatively specific and narrow scope. In which specific aspects of current LLM can this technology offer improvements? 2.This paper includes three toy experiments to verify the validity of the theory. Although multiple metrics are proposed for comparison, the experimental design is relatively simple. The last experiment on a real dataset is insufficient. The specific concerns are as follows: (1)Are the evaluation metrics sufficient? (2)Does the baseline methodology comprehensively cover cutting-edge technologies in the field? Is the comparison fair? (3)Could you elaborate on the computational cost of the technique proposed in this paper?

Questions for authors

The specific concerns are as follows: (1)The evaluation metrics does not seem to be sufficient, is it possible to improve it? (2)Does the baseline methodology comprehensively cover cutting-edge technologies in the field? Is the comparison fair? (3)Could you elaborate on the computational cost of the technique proposed in this paper?

3 comments

We thank the reviewer for their feedback, and we are pleased that they consider our work a “breakthrough in previous theoretical work”. Below, we provide new experimental evidence and analysis to address the reviewer’s concerns on 1) evaluation metrics, 2) fairness of baselines and 3) computational cost.

This paper involves highly technical theoretical research, resulting in a relatively specific and narrow scope.

We agree that our method involves technical challenges in stochastic calculus. However, we hope to convince the reviewer that these technicalities extend the scope of the method, rather than restricting it. Specifically, our central Theorem 1 provides guarantees for approximating stochastic dynamics over data manifolds, which are widespread in machine learning. Current methods in generative modelling focus on approximating such data manifolds rather than constructing dynamical processes over these manifolds, or target marginal distributions rather than the paths of stochastic processes. IMDs explicitly aim to recover the underlying dynamics.

Our experiments demonstrate that producing accurate sample paths is essential for faithful representations of random processes on implicit data manifolds, which we believe to be useful in downstream tasks.

In which specific aspects of current LLM can this technology offer improvements?

This is an interesting direction for future work. We hope that IMDs will enable stochastic exploration and optimization of geometrically constrained LLM parameter spaces [1], without explicit access to the geometry.

Questions

Are the evaluation metrics sufficient?

The aim of our experiments was to validate that IMDs provide accurate probability paths, our key contribution (see above). Our metrics (path MMD, QV, manifold distance) are common in stochastic simulation literature to measure path-space errors.

We compute the metrics in all relevant experiments, except the COIL-20 experiment (Fig. 11), which included no geometric reference process. We have now addressed this by generating a new synthetic image dataset which was sampled from a known ground-truth process. See below.

The last experiment on a real dataset is insufficient. Does the baseline methodology comprehensively cover cutting-edge technologies in the field? Is the comparison fair?

We thank the reviewer for the comment, which prompted us to strengthen our experiment on the image dataset. Let us remind the reviewer that one specific advantage of IMDs lies in approximating stochastic processes over implicit manifolds.

We now extended the COIL-20 experiment (Sec. 7.4, line 478 in the updated manuscript) by generating two-dimensional rotations of a 3D synthetic duck through a Brownian motion over the $S^1\times S^1 $ manifold, and rendering them into 32x32 pixel images (Figs. 6 in Sec. 7.4 and Fig. 9 in Appx. F.5.1.). We found that IMDs approximate the ground truth Brownian motion as well as the baseline graph-based random walk, measured by path MMD and QV (Fig. 10 in Appx F.5.1). However, perhaps more importantly, while the baseline graph process hops between data points, IMDs interpolate between data points consistently with the manifold geometry (see Fig. 10, Appx. F.5.1.).

IMDs achieve competitive performance against recent manifold exploration methods at substantially lower computational cost, while additionally providing path-space convergence guarantees. See next comment.

Could you elaborate on the computational cost of the technique proposed in this paper?

Thank you for raising this point. The computational cost of IMDs arises from an upfront cost of operator estimation followed by numerical simulation. The operator estimation uses proximity graphs, incurring a computational cost of O(N2D)O(N^2D). Numerical simulation uses low-dimensional neighbor searches (Sec. 7.1, 7.2, 7.3) using KD-trees. In the updated image experiment (Sec. 7.4), we avoid full D×DD\times D matrix operations by retaining the first rr eigendirections, giving a per-step cost of O(ND+krD)O(ND + krD).

In the new Table 4 in Section 7.3, we compare IMDs with recent manifold exploration methods [2,3]. We observed that IMDs provide comparable exploration metrics to [2] with a 200x speedup on CPU (15x speedup on GPU), and perform better than [3] across all metrics.

[1] Xie et al., mHC: Manifold-Constrained Hyper-Connections

[2] Sun et al. Geometry-Aware Generative Autoencoders for Warped Riemannian Metric Learning and Generative Modeling on Data Manifolds

[3] De Santi et al., Provable Maximum Entropy Manifold Exploration via Diffusion Models

AuthorsOpenReview ↗

Followup on our rebuttal

Dear Reviewer zy8W, we thank you again for your time and constructive feedback.

We have addressed your concerns with an extended experiment on the image dataset (including a ground truth and comparable baseline process), as well as a clarification on the used metrics and computational cost. Since the discussion period will soon be coming to an end, we would appreciate it if you could consider our rebuttal and improved manuscript and take these changes into account in your assessment.

We would be happy to respond to any further questions or concerns you may have.

Thanks for the reply, my concerns have been addressed. I would like to suggest authors to improve the presentation substantially despite I am keeping the positive rate for the submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.