pith. sign in

arxiv: 2605.00250 · v1 · submitted 2026-04-30 · 📊 stat.ML · cs.CV· cs.LG

Information-geometric adaptive sampling for graph diffusion

Yuhui Lu , Wenjing Liu , Kun Zhan This is my paper

Pith reviewed 2026-05-09 19:32 UTC · model grok-4.3

classification 📊 stat.ML cs.CVcs.LG
keywords graph diffusionadaptive samplinginformation geometryFisher-Rao metricDrift Variation Scoremolecule generationsocial network generationdiffusion models
0
0 comments X

The pith

Enforcing constant informational speed on the statistical manifold improves graph diffusion sampling quality.

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

Standard diffusion models for graphs advance with uniform time steps that ignore uneven rates of distributional change. This paper recasts the sampling trajectory as a curve on a Riemannian manifold and adopts the Fisher-Rao metric as the intrinsic measure of distance between successive distributions. From this geometry the authors derive the Drift Variation Score, which tracks the local rate of change and automatically adjusts step sizes so that informational speed stays constant. The resulting equal-arc-length discretization makes every step contribute the same amount of distributional progress. Experiments on molecule and social-network generation show gains in structural fidelity together with higher sampling efficiency.

Core claim

By treating diffusion sampling as motion along a parametric curve on the probability simplex equipped with the Fisher-Rao metric, the Drift Variation Score solver enforces constant informational speed, automatically producing an equal-arc-length discretization in which each step adds the same amount of distributional information and thereby improves the fidelity and efficiency of generated graphs.

What carries the argument

Drift Variation Score (DVS), the geometry-aware scalar derived from the Fisher-Rao metric that quantifies instantaneous distributional change rate and drives adaptive step sizes to hold informational speed fixed.

If this is right

  • Each discretization step contributes equally to information accumulation along the trajectory.
  • The sampling path maintains a uniform rate of distributional change measured in the Fisher-Rao sense.
  • Generated graphs exhibit higher structural fidelity on molecule and social-network tasks.
  • Fewer total steps are needed to reach a given level of distributional progress, raising sampling efficiency.

Where Pith is reading between the lines

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

  • The same constant-speed rule could be tested on image or text diffusion models whose distributions also evolve at varying rates.
  • Computing DVS may offer a principled substitute for heuristic adaptive schedulers used in other sampling algorithms.
  • On very large graphs the cost of estimating the Fisher-Rao-based score at each step could become a practical bottleneck worth measuring.

Load-bearing premise

The Fisher-Rao metric supplies the correct intrinsic distance for measuring how probability distributions evolve during graph diffusion, and keeping speed constant under that metric improves final sample quality.

What would settle it

If side-by-side runs of uniform time-stepping and DVS adaptive sampling on the same molecule and network benchmarks produce statistically indistinguishable scores on structural metrics such as validity, uniqueness, and MMD, the claimed benefit of constant informational speed would be falsified.

Figures

Figures reproduced from arXiv: 2605.00250 by Kun Zhan, Wenjing Liu, Yuhui Lu.

Figure 1
Figure 1. Figure 1: Conceptual illustration of DVS-driven adaptive sam￾pling on the statistical manifold M. We compare the standard fixed-step discretization (gray circles) with our DVS sampler (red stars). Fixed-step methods fail to account for the non-uniform dynamics of the sampling trajectory, leading to inconsistent in￾formational progress. In contrast, our DVS sampler maintains a constant informational distance per step… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of the information-geometric increment (∆s 2 ) on QM9 using GRUM. We compare the information progress of the fixed-step Euler-Maruyama sampler and the DVS￾driven Euler-Maruyama sampler. For the fixed-step sampler, the information increment remains small at early stages and increases sharply near the end, whereas the adaptive sampler maintains an approximately equal arc-length progression througho… view at source ↗
Figure 4
Figure 4. Figure 4: Numerical scale analysis of information variation components on QM9 using the GruM model. The drift component Vf (red solid line) represents the data-dependent variation of the transition distribution, which dominates the noise component Vg (blue dashed line) by over six orders of magnitude on average. The explosive growth of Vf near t = 1 highlights the intense structural transitions (crystallization) tha… view at source ↗
Figure 5
Figure 5. Figure 5: Visualizations of generated molecules on QM9 (top row) and ZINC250k (bottom row). The samples are generated using the GruM model with the DVS-Euler–Maruyama sampler. The top row shows organic molecules from QM9 with realistic geometric configurations. The bottom row displays complex drug-like molecules from ZINC250k, successfully capturing multi-ring systems and diverse heteroatoms (e.g., O, N, S, Cl) [PI… view at source ↗
Figure 6
Figure 6. Figure 6: Snapshots of structural evolution on the Planar dataset. The top row represents the standard fixed-step Euler sampler, while the bottom row represents our DVS-Euler–Maruyama sampler. As t increases, the DVS-driven sampler resolves the sparse planar skeleton more cleanly, especially in the late stages (t ≥ 0.6) [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Snapshots of structural evolution on the SBM dataset. The top row (fixed-step) and bottom row (DVS-Euler–Maruyama) both start from a dense noise state. However, the DVS-driven sampler achieves clearer community separation and fewer spurious inter-cluster edges by the final step (t = 1.0). 17 [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
read the original abstract

Standard diffusion models for graph generation typically rely on uniform time-stepping, an approach that overlooks the non-homogeneous dynamics of distributional evolution on complex manifolds. In this paper, we present an information-geometric framework that reinterprets the diffusion sampling trajectory as a parametric curve on a Riemannian manifold. Our key observation is that the Fisher-Rao metric provides a principled measure of the intrinsic distance. By analyzing this metric, we derive the Drift Variation Score (DVS), a geometry-aware indicator that quantifies the instantaneous rate of distributional change. Unlike prior heuristic-based adaptive samplers, our DVS solver enforces a constant informational speed on the statistical manifold, automatically maintaining a uniform rate of distributional change along the sampling trajectory. This equal arc-length strategy ensures that each discretization step contributes equally to the information speed. Theoretical analysis verifies that DVS characterizes the local stiffness of the sampling dynamics in the Fisher-Rao sense. Experimental results on molecule and social network generation show that DVS significantly improves structural fidelity and sampling efficiency. Code is at https://github.com/kunzhan/DVS

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper proposes an information-geometric adaptive sampling method for graph diffusion models. It reinterprets the diffusion sampling trajectory as a parametric curve on a Riemannian manifold equipped with the Fisher-Rao metric, derives a Drift Variation Score (DVS) that quantifies the instantaneous rate of distributional change, and uses a DVS solver to enforce constant informational speed (equal arc-length) along the trajectory. The authors claim this yields uniform per-step information contribution, characterizes local stiffness of the dynamics, and improves structural fidelity and sampling efficiency over uniform time-stepping, as demonstrated on molecule and social-network generation tasks.

Significance. If the central derivation is sound and the Fisher-Rao geometry is appropriate, the work would supply a principled, geometry-aware alternative to heuristic adaptive samplers in diffusion models. The equal-arc-length strategy and code release would be concrete strengths supporting reproducibility and potential downstream use in structured data generation. However, the significance is limited by the unresolved question of whether the probability-simplex Fisher-Rao metric correctly captures distributional change for discrete or relaxed graph states.

major comments (2)
  1. Abstract: the claim that DVS 'enforces a constant informational speed on the statistical manifold' and that 'each discretization step contributes equally to the information speed' presupposes that the diffusion trajectory is a smooth curve on the probability simplex equipped with the Fisher-Rao metric. For graph diffusion the state space consists of discrete graphs (or adjacency-matrix relaxations), so the relevant manifold and its tangent space are not obviously the simplex; without an explicit embedding or relaxation that preserves the Riemannian structure and differentiability of the SDE, the constant-speed property does not necessarily translate into uniform contribution to actual distributional evolution on graphs.
  2. Abstract (theoretical analysis paragraph): the statement that 'theoretical analysis verifies that DVS characterizes the local stiffness of the sampling dynamics in the Fisher-Rao sense' is presented without visible derivation steps, explicit equations relating DVS to the metric tensor, or a proof that the resulting adaptive step sizes remain well-defined when the underlying graph distribution is discrete. This absence makes it impossible to check whether the stiffness characterization is non-circular or reduces to a re-statement of the metric definition.
minor comments (2)
  1. Abstract: quantitative improvements, error bars, ablation tables, and statistical significance tests for the reported gains in fidelity and efficiency are not mentioned, which weakens the experimental claim.
  2. The manuscript would benefit from a clear statement of the precise manifold and coordinate chart used to embed graph states into the probability simplex before applying the Fisher-Rao metric.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the geometric foundations of our approach. We respond to each major comment below.

read point-by-point responses

  1. Referee: Abstract: the claim that DVS 'enforces a constant informational speed on the statistical manifold' and that 'each discretization step contributes equally to the information speed' presupposes that the diffusion trajectory is a smooth curve on the probability simplex equipped with the Fisher-Rao metric. For graph diffusion the state space consists of discrete graphs (or adjacency-matrix relaxations), so the relevant manifold and its tangent space are not obviously the simplex; without an explicit embedding or relaxation that preserves the Riemannian structure and differentiability of the SDE, the constant-speed property does not necessarily translate into uniform contribution to actual distributional evolution on graphs.

    Authors: We appreciate the referee highlighting the need for explicit manifold details. Our framework employs a continuous relaxation of adjacency matrices with entries in [0,1], embedding the states into a space where the probability simplex and Fisher-Rao metric apply directly. The SDE is defined on these relaxed states, yielding a differentiable trajectory on the Riemannian manifold. The constant-speed enforcement via DVS thus ensures uniform distributional change in the relaxed representation, which approximates the discrete graph dynamics. We will revise the abstract to state this relaxation explicitly and note how it preserves the required geometric and differentiability properties. revision: yes

  2. Referee: Abstract (theoretical analysis paragraph): the statement that 'theoretical analysis verifies that DVS characterizes the local stiffness of the sampling dynamics in the Fisher-Rao sense' is presented without visible derivation steps, explicit equations relating DVS to the metric tensor, or a proof that the resulting adaptive step sizes remain well-defined when the underlying graph distribution is discrete. This absence makes it impossible to check whether the stiffness characterization is non-circular or reduces to a re-statement of the metric definition.

    Authors: We agree the abstract is overly concise on the theory. The manuscript derives DVS as the Fisher-Rao norm of the SDE drift vector, DVS = sqrt(g_{ij} mu^i mu^j), where g is the metric tensor; this directly quantifies instantaneous distributional speed and identifies stiffness as high-norm regions. Adaptive steps are solved to maintain constant integrated DVS (equal arc length), which is well-defined under the continuous relaxation. We will update the abstract to reference this equation and direct readers to the theoretical section containing the full derivation and well-definedness argument for the relaxed setting. revision: yes

Circularity Check

0 steps flagged

No significant circularity; DVS derived from Fisher-Rao metric as first-principles construction

full rationale

The paper presents the Drift Variation Score (DVS) as derived from analysis of the Fisher-Rao metric on the statistical manifold, with the constant informational speed property following directly from the equal arc-length strategy under that metric. No load-bearing step reduces by the paper's own equations to a fitted parameter or self-referential definition; the theoretical verification that DVS characterizes local stiffness is stated as a consequence of the Riemannian geometry rather than an input. Experimental results on molecule and social network generation are reported separately and do not appear to be used to define or tune the core DVS equations. The derivation chain is self-contained against the stated geometric assumptions, with no self-citation load-bearing or renaming of known results as new derivations.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the appropriateness of the Fisher-Rao metric for measuring distributional change during discrete graph diffusion steps and on the assumption that equal arc-length discretization under this metric yields better samples; no free parameters or invented entities are mentioned in the abstract.

axioms (1)
  • domain assumption The Fisher-Rao metric supplies the intrinsic distance on the space of probability distributions over graphs.
    Invoked when the paper states that the metric provides a principled measure of intrinsic distance and derives DVS from it.

pith-pipeline@v0.9.0 · 5485 in / 1385 out tokens · 23481 ms · 2026-05-09T19:32:40.951730+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages

  1. [1]

    Information geometry and its applications: Survey

    Amari, S.-i. Information geometry and its applications: Survey. InGSI, pp. 3, 2013

    work page 2013

  2. [2]

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

    work page 1982

  3. [3]

    Ascher, U. M. and Petzold, L. R.Computer methods for ordinary differential equations and differential-algebraic equations. SIAM, 1998

    work page 1998

  4. [4]

    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. InICLR, 2022

    work page 2022

  5. [5]

    Equiv- ariant energy-guided SDE for inverse molecular design

    Bao, F., Zhao, M., Hao, Z., Li, P., Li, C., and Zhu, J. Equiv- ariant energy-guided SDE for inverse molecular design. InICLR, 2023

    work page 2023

  6. [6]

    N.Statistical decision rules and optimal infer- ence

    Cencov, N. N.Statistical decision rules and optimal infer- ence. Number 53. American Mathematical Soc., 2000

    work page 2000

  7. [7]

    On the trajectory regularity of ODE-based diffusion sampling

    Chen, D., Zhou, Z., Wang, C., Shen, C., and Lyu, S. On the trajectory regularity of ODE-based diffusion sampling. InICML, volume 41, pp. 7905–7934, 2024

    work page 2024

  8. [8]

    Chen, T., Liu, G., and Theodorou, E. A. Likelihood training of Schr ¨odinger bridge using forward-backward SDEs theory. InICLR, 2022

    work page 2022

  9. [9]

    and De Grave, K

    Costa, F. and De Grave, K. Fast neighborhood subgraph pairwise distance kernel. InICML, pp. 255–262, 2010. De Bortoli, V ., Thornton, J., Heng, J., and Doucet, A. Diffu- sion Schr¨odinger bridge with applications to score-based generative modeling. InNeurIPS, volume 34, pp. 17695– 17709, 2021

    work page 2010

  10. [10]

    and Calderhead, B

    Girolami, M. and Calderhead, B. Riemann manifold Langevin and Hamiltonian Monte Carlo methods.Jour- nal of the Royal Statistical Society Series B: Statistical Methodology, pp. 123–214, 2011

    work page 2011

  11. [11]

    Graphite: Iterative generative modeling of graphs

    Grover, A., Zweig, A., and Ermon, S. Graphite: Iterative generative modeling of graphs. InICML, pp. 2434–2444, 2019

    work page 2019

  12. [12]

    Denoising diffusion proba- bilistic models

    Ho, J., Jain, A., and Abbeel, P. Denoising diffusion proba- bilistic models. InNeurIPS, volume 33, pp. 6840–6851, 2020

    work page 2020

  13. [13]

    W., Laskey, K

    Holland, P. W., Laskey, K. B., and Leinhardt, S. Stochastic blockmodels: First steps.Social Networks, 5(2):109–137, 1983

    work page 1983

  14. [14]

    G., Vignac, C., and Welling, M

    Hoogeboom, E., Satorras, V . G., Vignac, C., and Welling, M. Equivariant diffusion for molecule generation in 3D. InICML, pp. 8867–8887, 2022

    work page 2022

  15. [15]

    J., Sterling, T., Mysinger, M

    Irwin, J. J., Sterling, T., Mysinger, M. M., Bolstad, E. S., and Coleman, R. G. ZINC: a free tool to discover chem- istry for biology.Journal of Chemical Information and Modeling, 52(7):1757–1768, 2012

    work page 2012

  16. [16]

    Jo, J., Lee, S., and Hwang, S. J. Score-based generative mod- eling of graphs via the system of stochastic differential equations. InICML, pp. 10362–10383, 2022

    work page 2022

  17. [17]

    Jo, J., Kim, D., and Hwang, S. J. Graph generation with diffusion mixture. InICML, volume 41, pp. 22371–22405, 2024

    work page 2024

  18. [18]

    Kahouli, K., Hessmann, S. S. P., M¨uller, K.-R., Nakajima, S., Gugler, S., and Gebauer, N. W. A. Molecular relax- ation by reverse diffusion with time step prediction.Ma- chine Learning: Science and Technology, 5(3):035038, 2024

    work page 2024

  19. [19]

    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. In NeurIPS, 2022

    work page 2022

  20. [20]

    and Ye, J

    Kim, B. and Ye, J. C. Denoising MCMC for accelerating diffusion-based generative models. InICML, pp. 16955– 16977, 2023

    work page 2023

  21. [21]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. InICLR, 2017

    work page 2017

  22. [22]

    and Ping, W

    Kong, Z. and Ping, W. On fast sampling of diffusion proba- bilistic models. InICML Workshop INNF, 2021

    work page 2021

  23. [23]

    Langevin, P. et al. Sur la th´eorie du mouvement brownien. CR Acad. Sci. Paris, 146(530-533):530, 1908

    work page 1908

  24. [24]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. InICML, pp. 1207–1216, Stanford, CA, 2000. Morgan Kaufmann

    work page 2000

  25. [25]

    Alleviating exposure bias in diffusion models through sampling with shifted time steps

    Li, M., Qu, T., Yao, R., Sun, W., and Moens, M. Alleviating exposure bias in diffusion models through sampling with shifted time steps. InICLR, 2024

    work page 2024

  26. [26]

    Fisher-Rao metric, geometry, and complexity of neural networks

    Liang, T., Poggio, T., Rakhlin, A., and Stokes, J. Fisher-Rao metric, geometry, and complexity of neural networks. In AISTATS, pp. 888–896, 2019

    work page 2019

  27. [27]

    I 2SB: Image-to-image Schr¨odinger bridge

    Nie, W., and Anandkumar, A. I 2SB: Image-to-image Schr¨odinger bridge. InICML, pp. 22042–22062, 2023. 9 Information-Geometric Adaptive Sampling for Graph Diffusion

    work page 2023

  28. [28]

    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. InNeurIPS, vol- ume 35, pp. 5775–5787, 2022

    work page 2022

  29. [29]

    GraphDF: A discrete flow model for molecular graph generation

    Luo, Y ., Yan, K., and Ji, S. GraphDF: A discrete flow model for molecular graph generation. InICML, pp. 7192–7203, 2021

    work page 2021

  30. [30]

    Spectre: Spectral conditioning helps to overcome the expressivity limits of one-shot graph generators

    Martinkus, K., Loukas, A., Perraudin, N., and Wattenhofer, R. Spectre: Spectral conditioning helps to overcome the expressivity limits of one-shot graph generators. InICML, pp. 15159–15179, 2022

    work page 2022

  31. [31]

    Continuous markov processes and stochastic equations.Rendiconti del Circolo Matematico di Palermo, 4(1):48–90, 1955

    Maruyama, G. Continuous markov processes and stochastic equations.Rendiconti del Circolo Matematico di Palermo, 4(1):48–90, 1955

    work page 1955

  32. [32]

    Constraint Ornstein-Uhlenbeck bridges.Jour- nal of Mathematical Physics, 58(9), 2017

    Mazzolo, A. Constraint Ornstein-Uhlenbeck bridges.Jour- nal of Mathematical Physics, 58(9), 2017

    work page 2017

  33. [33]

    Permutation invariant graph generation via score-based generative modeling

    Niu, C., Song, Y ., Song, J., Zhao, S., Grover, A., and Ermon, S. Permutation invariant graph generation via score-based generative modeling. InAISTATS, pp. 4474–4484, 2020

    work page 2020

  34. [34]

    Jump your steps: Optimizing sampling schedule of discrete diffusion models

    Park, Y ., Lai, C., Hayakawa, S., Takida, Y ., and Mitsufuji, Y . Jump your steps: Optimizing sampling schedule of discrete diffusion models. InICLR, volume 13, pp. 96272– 96300, 2025

    work page 2025

  35. [35]

    Fr ´echet chemnet distance: a metric for generative models for molecules in drug discovery.Jour- nal of Chemical Information and Modeling, 58(9):1736– 1741, 2018

    Klambauer, G. Fr ´echet chemnet distance: a metric for generative models for molecules in drug discovery.Jour- nal of Chemical Information and Modeling, 58(9):1736– 1741, 2018

    work page 2018

  36. [36]

    DeFoG: Discrete flow matching for graph generation

    Qin, Y ., Madeira, M., Thanou, D., and Frossard, P. DeFoG: Discrete flow matching for graph generation. InICML, volume 42, pp. 50269–50326, 2025

    work page 2025

  37. [37]

    O., Rupp, M., and V on Lilienfeld, O

    Ramakrishnan, R., Dral, P. O., Rupp, M., and V on Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules.Scientific Data, 1(1):1–7, 2014

    work page 2014

  38. [38]

    Align your steps: Optimizing sampling schedules in diffusion models

    Sabour, A., Fidler, S., and Kreis, K. Align your steps: Optimizing sampling schedules in diffusion models. In ICML, volume 41, pp. 42947–42975, 2024

    work page 2024

  39. [39]

    Collective classification in network data.AI Magazine, 29(3):93–93, 2008

    Sen, P., Namata, G., Bilgic, M., Getoor, L., Galligher, B., and Eliassi-Rad, T. Collective classification in network data.AI Magazine, 29(3):93–93, 2008

    work page 2008

  40. [40]

    GraphAF: a flow-based autoregressive model for molecu- lar graph generation

    Shi, C., Xu, M., Zhu, Z., Zhang, W., Zhang, M., and Tang, J. GraphAF: a flow-based autoregressive model for molecu- lar graph generation. InICLR, 2020

    work page 2020

  41. [41]

    and Lai, H

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

    work page arXiv 2024

  42. [42]

    and Ermon, S

    Song, Y . and Ermon, S. Generative modeling by estimating gradients of the data distribution. InNeurIPS, volume 32, 2019

    work page 2019

  43. [43]

    V ., and Niepert, M

    Tong, V ., Hoang, D., Liu, A., den Broeck, G. V ., and Niepert, M. Learning to discretize denoising diffusion ODEs. In ICLR, volume 13, pp. 47244–47282, 2025

    work page 2025

  44. [44]

    Uhlenbeck, G. E. and Ornstein, L. S. On the theory of the Brownian motion.Physical Review, pp. 823, 1930

    work page 1930

  45. [45]

    DiGress: Discrete denoising diffusion for graph generation

    Vignac, C., Krawczuk, I., Siraudin, A., Wang, B., Cevher, V ., and Frossard, P. DiGress: Discrete denoising diffusion for graph generation. InICLR, 2023

    work page 2023

  46. [46]

    Learning fast samplers for diffusion models by differentiating through sample quality

    Watson, D., Chan, W., Ho, J., and Norouzi, M. Learning fast samplers for diffusion models by differentiating through sample quality. InICLR, 2022

    work page 2022

  47. [47]

    GeoDiff: A geometric diffusion model for molecular conformation generation

    Xu, M., Yu, L., Song, Y ., Shi, C., Ermon, S., and Tang, J. GeoDiff: A geometric diffusion model for molecular conformation generation. InICLR, 2022

    work page 2022

  48. [48]

    Accelerating diffusion sampling with optimized time steps

    Xue, S., Liu, Z., Chen, F., Zhang, S., Hu, T., Xie, E., and Li, Z. Accelerating diffusion sampling with optimized time steps. InCVPR, pp. 8292–8301, 2024

    work page 2024

  49. [49]

    and Chen, Y

    Zhang, Q. and Chen, Y . Fast sampling of diffusion models with exponential integrator. InICLR, 2023

    work page 2023

  50. [50]

    GraphVite: A high- performance CPU-GPU hybrid system for node embed- ding

    Zhu, Z., Xu, S., Tang, J., and Qu, M. GraphVite: A high- performance CPU-GPU hybrid system for node embed- ding. InWWW, pp. 2494–2504, 2019. 10 Information-Geometric Adaptive Sampling for Graph Diffusion Appendix OrganizationThe Appendix is organized as follows: InSection A, we provide thedetailed mathematical derivations for the information-geometric fra...

    work page 2019