acceptodds
Under review as a conference paper at ICLR 2027

Graph Residual Conjugate Diffusion: SNR-Equalized Heat Flow for Graph Signals

Abstract

Diffusion models generate data by reversing a forward corruption process that typically approaches a simple Gaussian prior. Recent work has extended this framework to signals supported on fixed graphs, e.g., road-network traffic and sensor-network measurements. Many graph signals have nonuniform spectral energy across graph frequencies, whereas isotropic corruption adds the same conditional noise variance to every graph-frequency mode. Driving all modes to near-zero terminal signal-to-noise ratio (SNR) requires strong corruption, which increases the noise range that must be covered under a fixed sampling budget. We introduce Graph Residual Conjugate Diffusion (GRCD), which replaces the shared clock of graph heat diffusion with a mode-dependent clock that gives every graph-Fourier mode the same conditional signal-to-noise ratio. GRCD fits a zero-mean graph-spectral Gaussian reference on the training split and, rather than driving the corruption to near-zero terminal SNR, stops at a finite terminal SNR at which the propagated reference still carries the fitted spectral variances, narrowing the SNR range that sampling must cover. The Gaussian component has an exact modewise propagator in the probability-flow ODE, so sampling advances it analytically and integrates only the learned residual score numerically. We evaluate GRCD on five graph-signal settings (the METR-LA traffic and Molene weather datasets, and three synthetic settings from stochastic block models) against seven comparators under a matched protocol: Graph-Aware Diffusion (GAD), EDM adapted to the graph backbone, two graph adaptations of Whitened Score Diffusion (WSD), and three preconditioning controls. At four function evaluations (NFEs), GRCD lowers averaged maximum mean discrepancy (aMMD) by – over the best comparator on all five settings, reaching on METR-LA, where it clears an aMMD target with % less sampling wall-clock time than the cheapest comparator that reaches it. Within GRCD, fitting the terminal reference reduces aMMD by factors of – at finite terminal SNR, while the corresponding factors shrink to – near zero, matching the analytic limit in which the propagated reference loses dependence on the fitted covariance.

Primary area:
generative models
Keywords:
diffusion models, graph diffusion, graph signal processing, few-step sampling, probability flow ODE, spectral graph theory

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3d8LOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces GRCD for generating signals on a fixed graph. Its graph-dependent forward process becomes isotropic variance-exploding diffusion after an invertible change of coordinates. A Gaussian reference is fitted to the training data in the graph-Fourier basis and propagated to a finite noise level to initialize sampling. The score is decomposed into an analytic Gaussian component and a learned residual, allowing the Gaussian part of the probability-flow equation to be integrated exactly. The clearest result is that fitting the terminal reference improves generation at small sampling budgets under the reported metric. The construction is coherent, and the ablations are informative. My main reservation is whether the experiments isolate the additional benefit of the graph-dependent representation once the fitted reference and finite noise range are shared. I would also like more direct evidence that the model recovers the non-Gaussian structure deliberately included in the synthetic data.

Critical strengths

The main derivations fit together clearly. The graph-filter transformation, residual denoising target, and Gaussian propagator appear internally consistent. The paper also carefully distinguishes the propagated Gaussian reference from the true forward marginal.

The terminal-reference comparison tests a specific mechanism. Section 5.2 holds the score model fixed while changing the sampling initialization. The benefit of fitting the reference is large at finite terminal signal-to-noise ratio and nearly disappears near zero terminal SNR, consistent with the analytic limit. In Table 2, fitting the reference reduces METR-LA averaged maximum mean discrepancy from 0.3722 to 0.0513 at 16 network function evaluations.

The experiments include several useful controls. The Gaussian controls, DPM-Solver++ comparison, training-only covariance fitting, validation-based grid selection, uncertainty estimates, and wall-clock measurements help interpret the results. The authors also report cases where other methods perform better at larger sampling budgets.

Critical weaknesses

1. The practical effect of the terminal approximation remains unclear.

Appendix A.2.3 explicitly acknowledges that the propagated Gaussian reference generally differs from the true forward marginal. The remaining question is how much the approximation affects generation in the finite-SNR regime. Even with the exact score, applying the same invertible probability-flow map over a finite interval to two different initial distributions leaves different output distributions. However, the observation alone does not quantify the effect on aMMD or determine the behavior of the implemented finite-step sampler.

Section 5.2 provides useful evidence that fitting the reference improves initialization relative to a scalar Gaussian. A complementary diagnostic would measure the remaining discrepancy from the true terminal distribution. On the synthetic mixture, the authors could compare the fitted terminal reference with samples obtained by forward-corrupting fresh independent draws from the known data generator. Running both initializations through the same trained model, solver, grid, and NFE would test how sensitive generation is to the terminal approximation. The diagnostic draws should be separate from the final evaluation samples.

The reported low-budget gains remain useful. The proposed check would help assess the distributional cost of the approximation, especially alongside the direct test of the planted mixture requested below.

2. The additional benefit of the graph-dependent representation needs a more direct comparison.

Proposition 1 expresses the graph-dependent process as isotropic VE diffusion under an invertible transformation. The transformation can improve learning or numerical conditioning, and Appendix A.1.2 explains the conditional-SNR result within the graph heat-diffusion family. My concern is whether the experiments separate the benefit of the graph transformation from that of the fitted Gaussian reference and finite noise range.

Let qq denote the scalar noise level in the conjugate isotropic coordinates. The existing isotropic and preconditioning controls use the noise range q∈[0.02,32.64]q \in [0.02, 32.64], compared with q∈[0.04,2]q \in [0.04, 2] for finite-SNR GRCD. The whitened-score graph-prior control in Appendix A.6.4 shares the fitted covariance and finite noise regime, but retains the conjugate graph forward process. The existing comparisons therefore address parts of the question without fully isolating the graph transformation.

I would like a control that independently trains the same graph backbone directly on isotropically corrupted node signals, using the same fitted graph-spectral Gaussian reference and finite scalar noise range. The control should learn a residual relative to the Gaussian score under its own forward process. Each method should initialize from its corresponding propagated reference, with matched training and sampling budgets and the same protocol for choosing the solver and grid.

Wang and Vastola (2024), already cited by the paper, discuss propagated Gaussian initialization and learning a residual beyond a Gaussian score. A matched comparison would help identify the additional value of the graph-dependent representation. The current results show that the combined method works well under the chosen metric, but leave the size of that additional benefit unclear.

3. The synthetic evaluation does not directly test recovery of the planted non-Gaussian structure.

The reported aMMD compares the distributions of three scalar summaries: quadratic variation, spectral centroid, and degree correlation. These statistics measure useful properties of graph signals, but they do not directly test recovery of the planted bimodality. Appendix A.8.1 acknowledges the limitations of the metric; a direct check of the known synthetic mixture would make the evidence more complete.

The Gaussian controls provide meaningful evidence. In Table 11, GRCD achieves an aMMD of 0.0538 on METR-LA, compared with 0.1921 for the best direct Gaussian control. However, on the stochastic block model setting with spectral concentration parameter c=1c=1, the fitted graph-spectral Gaussian reaches 0.0535, close to GRCD's 0.0497. The small gap under aMMD leaves open whether GRCD recovers the bimodal coefficient distribution in that setting.

Please show the generated coefficient distribution along the Fiedler vector used by the synthetic data generator, with preprocessing handled consistently. A multivariate two-sample comparison on complete graph signals would also help assess joint structure beyond the three summaries. The direct mixture diagnostic is the more immediate request because the non-Gaussian feature is known by construction.

4. The runtime comparison with the strongest matched solver remains incomplete.

Section 5.4 already explains that the exponential-residual solver is competitive with DPM-Solver++ without consistently outperforming it. In Table 3, DPM-Solver++ 2M with power-law grid parameter ρ=3\rho=3 achieves an aMMD of 0.0523 at four actual NFE, compared with 0.0552 for the proposed solver. The proposed solver has lower means at the other reported budgets.

However, the matched DPM-Solver++ configuration is absent from Figure 2 and Table 12. Its measured runtime under the same model, initialization, and grid would help assess which solver offers the better quality-cost tradeoff. I view the missing runtime comparison as a secondary concern; it does not invalidate the reported speedup over the external baselines and preconditioning controls.

Questions for authors

Q1. On the synthetic mixture, how close is the fitted terminal reference to the true forward marginal? How do generated samples change when the initialization is replaced by forward-corrupted independent synthetic draws, with the trained model, solver, grid, and NFE held fixed?

Q2. How does GRCD compare with isotropic VE diffusion independently trained directly on node signals, using the same fitted Gaussian reference and scalar noise interval, with each terminal reference propagated under its respective forward process? Under matched training and sampling protocols, what additional benefit comes from the graph-transformed representation?

Q3. Does GRCD recover the planted bimodal coefficient distribution along the Fiedler vector used by the synthetic data generator? How do the methods compare under a multivariate two-sample measure on complete graph signals?

Q4. Can the DPM-Solver++ 2M configuration from Table 3, with power-law grid parameter ρ=3\rho=3, be included in Figure 2 and Table 12 with measured sampling runtime?

4 comments

We appreciate the time and care the reviewer has devoted to our work. Your comments have helped us refine our ideas and improve the presentation. We have incorporated your suggestions into the revised manuscript, with all changes highlighted in green for ease of review.

Reviewer: Q1.1 On the synthetic mixture, how close is the fitted terminal reference to the true forward marginal?

By "the synthetic mixture", we assume you mean our stochastic block model (SBM) setting, for which the true forward marginal can be computed because we generated this synthetic dataset. We drew 5,000 samples from the fitted terminal reference and 5,000 samples from the true forward marginal, and compared the two sample sets with the same aMMD the paper uses to compare the results. On SBM c = 1, kappa = 2, the distance between the fitted terminal reference and the true forward marginal is 0.0028 measured in aMMD, which is less than half the seed-to-seed standard deviation on this setting (±0.0062 for GRCD at NFE 4, this number is in the paper in Appendix: Additional Benchmark Results). They do differ in shape, because the true terminal law is a mixture with two peaks, and the bimodal component survives the forward process, while a Gaussian distribution has only one peak.

Reviewer: Q1.2 How do generated samples change when the initialization is replaced by forward-corrupted independent synthetic draws, with the trained model, solver, grid, and NFE held fixed?

We thank the reviewer for this valuable question. We conduct this experiment on SBM because its known synthetic generator allows us to draw independent samples from the true forward marginal. For real-world datasets such as METR-LA and Molene, the underlying data distributions are unknown, so initialization from the true forward marginal is not directly feasible. Our fitted Gaussian reference is what makes the method usable without access to the true data distribution. This experiment provides useful insight into the potential improvement from a more accurate terminal reference.

We performed the experiment as suggested by changing the initialization in the sampler. Normally it starts from a sample of our fitted Gaussian reference. We replaced that draw with fresh generator samples, from a stream independent of train, validation and test, pushed forward to the terminal time. Everything else stayed fixed. We then measured the final generated samples against the real test data, on 5,000 generated samples against 5,000 test data. At four network evaluations, aMMD falls from 0.0497 ± 0.0062 to 0.0261 ± 0.0037 (mean ± std over 3 seeds).

Reviewer: Q2. How does GRCD compare with isotropic VE diffusion independently trained directly on node signals, using the same fitted Gaussian reference and scalar noise interval, with each terminal reference propagated under its respective forward process?

We present a table of experimental results in the following. In this table, “Control” means ‘isotropic VE diffusion independently trained directly on node signals, using the same fitted Gaussian reference and scalar noise interval, with each terminal reference propagated under its respective forward process’.

GRCD vs. Control. METR-LA, κ=2\kappa = 2, converged, mean ± s.d. over 3 seeds. Lower is better.

MethodNFE 4NFE 8NFE 16NFE 64
GRCD, =30.0538 ± 0.00370.0503 ± 0.00280.0509 ± 0.00150.0529 ± 0.0028
Control, =30.0532 ± 0.00410.0506 ± 0.00610.0519 ± 0.00330.0541 ± 0.0046
GRCD, Heun0.1618 ± 0.01370.0622 ± 0.00280.0537 ± 0.00340.0531 ± 0.0033
Control, Heun0.2239 ± 0.01460.0678 ± 0.00500.0551 ± 0.00300.0545 ± 0.0033

Reviewer: Under matched training and sampling protocols, what additional benefit comes from the graph-transformed representation?

In the above table, row 1 and row 2 are very near, which is explainable in a mathematical sense. Rows 1 and 2 are close by construction. Proposition 1 shows that the two forward processes are the same process seen through an invertible rescaling, so with an exact score in continuous time they produce the same distribution. The Control is our own process with the graph shrinkage switched off, and it inherits the same closed-form reference and propagator. What the transformation provides is tractability for a graph-dependent corruption. Diagonalising the graph-dependent dynamics into independent spectral modes is what puts the forward transition, the propagated reference, the score and an exact modewise propagator in closed form for a single process, which is what graph heat diffusion lacks. That closed form is also what makes the mode-dependent SNR visible and controllable, which is why the noise scale is set by a terminal-SNR matching rule.

The Control considered here is distinct from the Isotropic VE baseline in Table 1 (in the paper). The paper’s Isotropic VE is a preconditioning baseline, which lacks the fitted reference, the residual-score parameterization and GRCD’s noise interval. The Control reported here supplies the isolation the question asks for: it is GRCD with only the graph transformation removed, which is why it reaches 0.0532 where the Table 1 baseline reaches 2.7581.

Our stated third contribution derives what happens to the propagated reference as the terminal noise grows, using the transformed representation.

As we can see in the last two table rows in the above table, with a second-order solver on a uniform grid, the graph-transformed process is ahead at every budget level and for every seed at NFE 4 to 16. And because seeds are shared it is lower in all nine seed-budget pairs at NFE 4, 8 and 16 (at NFE 16: 0.0573/0.0578, 0.0507/0.0519, 0.0532/0.0556).

Reviewer: Q3.1 Does GRCD recover the planted bimodal coefficient distribution along the Fiedler vector used by the synthetic data generator?

We interpret "recover" in this context as reproducing two distinct peaks. Indeed, in our experiments GRCD cannot reproduce this shape. The generator places two peaks near +3 and -3, along the Fiedler vector (the eigenvector of the graph Laplacian with the smallest nonzero eigenvalue, which separates the two communities). We project all samples onto it and count those in the gap (gap = 1.0 × component_sd = 0.9659, e.g., |x| < 0.9659, component_sd stands for component's standard deviation). The real data has 1.4% there, while GRCD has 24.0%± 1.1% (NFE 4, ρ=3\rho = 3, three seeds, 5,000 samples each). This is close to what a single Gaussian of the same width would give (about 24% for that interval), and the standard deviations agree (3.13 against 3.14), as expected because our reference fits the variance.

The mismatch is based mainly on the Gaussian starting distribution, which cannot have two peaks. This explains why aMMD, which relies on second-order statistics, is insensitive to this difference. We add this observation to the limitations in the revised submission. A non-Gaussian reference, such as a Gaussian mixture, is a natural extension, but it would require adapting the exact Gaussian step, which we leave for future work.

Reviewer: Q3.2 How do the methods compare under a multivariate two-sample measure on complete graph signals?

We thank the reviewer for this question as this led us to perform additional experiments, which we describe in the following. We evaluated all methods with the multivariate energy distance V-statistic (Baringhaus and Franz, 2004; Székely and Rizzo, 2013) on complete 32-node graph signals, with n=m=1500, on SBM c = 1. All methods are evaluated in the same space, using per-node z-score normalization with statistics estimated exclusively from the training split, averaged over three seeds. The full Ledoit–Wolf Gaussian is a non-learned control with no sampler, evaluated over three independent draws to estimate sampling variability. The reference row compares two disjoint samples of real data in a single evaluation. The empirical energy distance V-statistic is defined as

\\;-\\; \frac{1}{n^{2}}\sum_{i=1}^{n}\sum_{j=1}^{n}\lVert X_i - X_j\rVert_2 \\;-\\; \frac{1}{m^{2}}\sum_{i=1}^{m}\sum_{j=1}^{m}\lVert Y_i - Y_j\rVert_2 ,$$ | Method | NFE 4 | NFE 16 | NFE 64 | |---|---|---|---| | Two disjoint real-data samples (reference) | 0.0104 | 0.0104 | 0.0104 | | Full Ledoit–Wolf Gaussian (non-learned, 3 draws) | 0.0161 ± 0.0008 | 0.0161 ± 0.0008 | 0.0161 ± 0.0008 | | **GRCD, $\rho = 3$** | **0.0194 ± 0.0068** | **0.0174 ± 0.0013** | 0.0185 ± 0.0008 | | Mode-dependent clock only | 0.0495 ± 0.0017 | 0.0531 ± 0.0017 | 0.0530 ± 0.0017 | | WSD-style, graph prior | 0.4390 ± 0.1361 | 0.7640 ± 0.1223 | 0.2850 ± 0.0487 | | WSD-style, scalar | 0.4754 ± 0.1483 | 1.0755 ± 0.1497 | 0.4417 ± 0.0638 | | Isotropic VE | 46.23 ± 0.07 | 0.0484 ± 0.0020 | **0.0135 ± 0.0009** | | Conjugate precond. | 46.26 ± 0.09 | 0.1640 ± 0.0971 | 0.0915 ± 0.0637 | | Static precond. | 46.32 ± 0.10 | 0.0684 ± 0.0134 | 0.0375 ± 0.0158 | | EDM | 124.32 ± 0.24 | 0.1490 ± 0.0041 | **0.0132 ± 0.0009** | | GAD (matched) | 664.31 ± 3.41 | 52.02 ± 0.11 | 44.37 ± 0.10 | On this metric the non-learned Gaussian is as good as GRCD, so the learned residual brings no measurable gain on SBM c = 1. This matches the remark in the appendix that the fitted Gaussian reaches 0.0535 against GRCD’s 0.0497 at c = 1. Every GRCD variant lands between 0.0166 and 0.0194 at NFE 4, within one seed standard deviation of each other, and all are far below the comparison methods. By NFE 64, EDM and isotropic VE overtake it (0.0132 and 0.0135 against 0.0185), consistent with the reversal we already report on the SBM settings in the main paper. In conclusion, GRCD's advantage lies in the low-budget regime. The synthetic generator is largely Gaussian. In its graph-Fourier basis, 31 of the 32 coefficients are Gaussian, with bimodality only along the Fiedler coefficient. A Euclidean-distance measure on all 32 coordinates dilutes the mismatch in this single direction, so, like aMMD, it is insensitive to the planted structure. The Fiedler projection analysis performed in Q3.1 is a more sensitive test of that structure. References. L. Baringhaus and C. Franz. On a new multivariate two-sample test. Journal of Multivariate Analysis, 88(1):190-206, 2004. G. J. Szekely and M. L. Rizzo. Energy statistics: A class of statistics based on distances. Journal of Statistical Planning and Inference, 143:1249-1272, 2013.

Reviewer: Q4. Can the DPM-Solver++ 2M configuration from Table 3, with power-law grid parameter rho=3, be included in Figure 2 and Table 12 with measured sampling runtime?

Thank you for the suggestion. We have added “GRCD with DPM-Solver++ 2M (ρ=3\rho = 3)” to Figure 2 and Table 15 (in the revised manuscript, it was previously Table 12). It uses the same trained model and forward process, and only the solver differs. Runtimes are measured on the same machine as the other entries ((Apple M5 Pro, MPS backend).

At NFE 4 it costs 0.50 seconds per 1000 samples against 0.52 for our solver, and it is the more accurate of the two there (aMMD 0.0523 ± 0.0010 against 0.0552 ± 0.0024 over three seeds). This configuration was already reported in Table 4 in the revised manuscript.

We thank the reviewer for this valuable suggestion.

Reviewer_suvaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper addresses the problem of graph-signal generation. The proposed method, GRCD, consists of a score-based diffusion model, with two major modifications: a graph-frequency-dependent noising process that equalizes the conditional SNR across graph frequencies, and a finite-SNR terminal distribution based on a fitted Gaussian reference (i.e., instead of running the noising process until almost all information is destroyed, the authors stop at a point where some fundamental information is still preserved). GRCD demonstrates promising performance in low budget settings (low NFEs/sampling time). The paper provides the theoretical derivations of their formulation and the experimental details of their implementation, which (although the code is not provided for now; only released upon acceptance) seem sufficient to reproduce their results.

Critical strengths
  1. The paper is theoretically sound (to the best of my understanding);
  2. GRCD generally outperforms the explored baselines in terms of aMMD on the adopted benchmarks in low budget regimes (low NFEs and low wall-clock time).
Critical weaknesses

My main concern with this paper regards its presentation and its structure, at several levels:

  • It doesn't contextualize the paper in terms of the overall motivation, the problem addressed, its relevance, and what previous works left uncovered. The absence of a devoted and compact related work section where the related papers, problems addressed, and their approaches are put in perspective, does not help in this regard. Instead, related works are introduced abruptly in the middle of the narrative, significantly breaking the flow of the reader;
  • Specialized terminology is frequently introduced without providing sufficient conceptual background/intuition (e.g., "clock"), making it even more difficult for the reader to follow;
  • There is no main figure that explains the overall method. This could partially address the two points mentioned above;
  • Most of the results reported in the main paper concern a single dataset (METR-LA), with the exception of Figure 1 (left). Although there are more results on the remaining datasets in the Appendix, these are important to bring into the main paper, as they support the empirical claims about GRCD.

All of these make it very difficult for the reader to follow where the paper is heading. More importantly, the lack of an appropriate global perspective in the paper severely harms the reader's understanding of the significance of the results obtained.

Questions for authors

Clarification Requests

  1. There is no reference to Table 3 at any point in the paper. Can you confirm that these are the results discussed in Section 5.4?
  2. The evaluation of generative quality in the paper relies exclusively on aMMD, following previous works. This metric only averages MMD over three graph statistics, potentially allowing models to perform well without necessarily capturing the full data distribution. Would it be possible for the authors to report additional evaluation metrics to provide a broader assessment of the quality of the generated samples?
  3. The paper focuses on low budget regimes for iterative refinement methods, but fully ignores existing generative frameworks specifically designed for few-step generation (e.g., Consistency Models [1] or MeanFlow [2]). These methods should at least be discussed: is there any fundamental limitation preventing these methods from being applied in this setting? If not, why not evaluate them as baselines?

[1] - Consistency Models, Song et al., ICML 2023

[2] - Mean Flows for One-step Generative Modeling, Geng et al., NeurIPS 2025

6 comments

We appreciate the time and care the reviewer has devoted to our work. Your comments have helped us refine our ideas and improve the presentation. We have incorporated your suggestions into the revised manuscript, with all changes highlighted in green for ease of review.

Reviewer: There is no reference to Table 3 at any point in the paper. Can you confirm that these are the results discussed in Section 5.4?

Thank you for catching this oversight and we apologize for the missing reference. We confirm that Table 4: Solver comparison contains the results discussed in Section 5.4. We have now added the corresponding reference in the revised manuscript to make this more clear.

Reviewer: The evaluation of generative quality in the paper relies exclusively on aMMD, following previous works. This metric only averages MMD over three graph statistics, potentially allowing models to perform well without necessarily capturing the full data distribution. Would it be possible for the authors to report additional evaluation metrics to provide a broader assessment of the quality of the generated samples?

We additionally evaluate sliced Wasserstein distance (Rabin et al., 2011), averaging the one-dimensional 1-Wasserstein distance over 300 random unit projections. Unlike aMMD, which reduces each sample to three graph statistics, these projections operate on the full 207-dimensional signal. Sliced Wasserstein distance has also been used as a principal measure of sample quality in generative modeling (Karras et al., 2018).

At NFE 4, GRCD reaches 0.1839 against 0.4491 for the best comparison method (WSD-style, scalar), a factor of 2.4. The ablation without the residual score reaches 0.1770, a factor of 2.5. At NFE 64 GRCD still leads, at 0.1763 against 0.1860 for isotropic VE and 0.1862 for EDM. Among the GRCD variants the differences are within one standard deviation from NFE 8 onward, so this metric separates GRCD from the comparison methods but does not separate the GRCD ablations from one another.

We note explicitly that the margin here is much smaller than under aMMD, where the corresponding factor on METR-LA is about 23 (Section 5.5). The qualitative ranking is the same under both metrics, but the large factors quoted in the paper are specific to aMMD.

The evaluation now encompasses a broader range of complementary metrics. Taken together with the energy distance (Q3.2 from Reviewer 3d8L) and the Fiedler-projection recovery test (Q3.1 from Reviewer 3d8L), the evaluation now covers sliced Wasserstein over random projections of the full signal, energy distance on the raw samples, and a targeted test of a single graph-frequency coordinate.

Sliced Wasserstein distance on METR-LA.

κ=2\kappa = 2, converged, mean ± s.d. over three random seeds. Lower is better and bold denotes the best mean in each column. The same 300 projections, drawn once from a fixed seed, are used for every method, seed and budget, and are shared between the generated and reference samples.

MethodNFE 4NFE 8NFE 16NFE 64
EDM32.2509 ± 0.48821.5345 ± 0.05770.2660 ± 0.00530.1862 ± 0.0018
GAD142.5409 ± 0.126017.5713 ± 0.073813.6073 ± 0.073511.7474 ± 0.0799
WSD-style, scalar0.4491 ± 0.05800.3727 ± 0.04340.3382 ± 0.03480.2898 ± 0.0193
WSD-style, graph prior0.5012 ± 0.05730.3015 ± 0.03530.2574 ± 0.02840.2156 ± 0.0156
Isotropic VE12.4228 ± 0.09560.6655 ± 0.01480.2129 ± 0.00560.1860 ± 0.0028
Static precond.12.5519 ± 0.01380.6549 ± 0.01740.2100 ± 0.00500.1887 ± 0.0010
Conjugate precond.12.4734 ± 0.14540.6826 ± 0.02870.2124 ± 0.00030.1865 ± 0.0015
GRCD w/o residual score0.1770 ± 0.00080.1751 ± 0.00090.1756 ± 0.00100.1757 ± 0.0009
GRCD, ρ=1\rho = 10.2462 ± 0.03670.1805 ± 0.00480.1772 ± 0.00190.1767 ± 0.0012
GRCD, ρ=3\rho = 30.1839 ± 0.00370.1778 ± 0.00250.1767 ± 0.00010.1763 ± 0.0014

Reference.

J. Rabin, G. Peyré, J. Delon and M. Bernot. Wasserstein barycenter and its application to texture mixing. International Conference on Scale Space and Variational Methods in Computer Vision, pp. 435–446, 2011.

T. Karras, T. Aila, S. Laine and J. Lehtinen. Progressive growing of GANs for improved quality, stability, and variation. International Conference on Learning Representations (ICLR), 2018.

Reviewer: The paper focuses on low budget regimes for iterative refinement methods, but fully ignores existing generative frameworks specifically designed for few-step generation (e.g., Consistency Models [1] or MeanFlow [2]). These methods should at least be discussed: is there any fundamental limitation preventing these methods from being applied in this setting? If not, why not evaluate them as baselines? [1] - Consistency Models, Song et al., ICML 2023 [2] - Mean Flows for One-step Generative Modeling, Geng et al., NeurIPS 2025

There is no fundamental limitation and we did not ignore the mentioned frameworks on purpose. We trained both consistency models (CT) and MeanFlow on METR-LA and report their respective results below with a matched and an extended training budget. They differ in requiring a different training objective and their own tuning, whereas GRCD keeps the standard denoising objective and gains from the fitted starting distribution and the solver. Hence, these approaches are complementary, and combining them is future work.

As the tables show, with extended training, CT is better than GRCD at one and two network evaluations (aMMD 0.1258 and 0.0720 at NFE 1 and 2). From NFE 4 GRCD is best in aMMD, and comparable to CT in SWD at NFE 4 and better from NFE 8. CT degrades with more steps, and MeanFlow plateaus at a higher error. We have changed the paper accordingly to state that GRCD’s advantage starts at four network evaluations.

Table 1. Few-step baselines vs. GRCD, aMMD.

κ=2\kappa = 2, METR-LA. Mean ± standard deviation over three seeds. Lower is better. GRCD is trained to convergence with validation-based early stopping, which stops at 6,250 updates on average. The baselines are reported at a fixed 5,000-update budget and at an extended 40,000-update budget. NFE 1 is omitted because GRCD's exponential-residual sampler requires two network evaluations per step and has no entry there. At NFE 1 the one-step baselines reach 0.2585 ± 0.0465 (CT, 5,000), 0.1258 ± 0.0206 (CT, 40,000), 0.2472 ± 0.0255 (MeanFlow, 5,000) and 0.2652 ± 0.0485 (MeanFlow, 40,000). MeanFlow uses the authors' published method constants without domain-specific tuning.

MethodNFE 2NFE 4NFE 8NFE 16NFE 32NFE 64
CT, 5,000 updates0.2019 ± 0.03990.2051 ± 0.03620.2060 ± 0.03490.2742 ± 0.02890.5261 ± 0.01281.0461 ± 0.0525
CT, 40,000 updates0.0720 ± 0.00940.0667 ± 0.00900.0778 ± 0.01160.1524 ± 0.01760.3641 ± 0.01320.8124 ± 0.0234
MeanFlow, 5,000 updates0.2436 ± 0.05870.2133 ± 0.04330.2099 ± 0.03760.2053 ± 0.03240.2000 ± 0.02860.1962 ± 0.0264
MeanFlow, 40,000 updates0.2118 ± 0.01240.1692 ± 0.00710.1726 ± 0.00470.1762 ± 0.00480.1770 ± 0.00510.1768 ± 0.0053
GRCD, ρ=3\rho = 30.9041 ± 0.09450.0538 ± 0.00370.0503 ± 0.00280.0509 ± 0.00150.0540 ± 0.00310.0529 ± 0.0028

Table 2. Few-step baselines vs. GRCD, sliced Wasserstein distance.

κ=2\kappa = 2, METR-LA. Mean ± standard deviation over the same three seeds. Lower is better. Training budgets are as in Table 1. NFE 1 is omitted for the same reason. At NFE 1 the one-step baselines reach 0.1930 ± 0.0024 (CT, 5,000), 0.1880 ± 0.0018 (CT, 40,000), 0.3222 ± 0.0199 (MeanFlow, 5,000) and 0.3460 ± 0.0099 (MeanFlow, 40,000). MeanFlow uses the authors' published method constants without domain-specific tuning.

MethodNFE 2NFE 4NFE 8NFE 16NFE 32NFE 64
CT, 5,000 updates0.1967 ± 0.00100.2136 ± 0.00370.2499 ± 0.01060.3703 ± 0.04460.8044 ± 0.17642.0374 ± 0.4790
CT, 40,000 updates0.1805 ± 0.00130.1844 ± 0.00320.1965 ± 0.00480.2490 ± 0.01560.4819 ± 0.03661.2600 ± 0.1544
MeanFlow, 5,000 updates0.3477 ± 0.01630.3078 ± 0.00240.2953 ± 0.00260.2893 ± 0.00300.2854 ± 0.00250.2829 ± 0.0020
MeanFlow, 40,000 updates0.3419 ± 0.00570.3038 ± 0.00380.2920 ± 0.00410.2866 ± 0.00540.2835 ± 0.00650.2817 ± 0.0071
GRCD, ρ=3\rho = 31.0975 ± 0.28970.1839 ± 0.00370.1778 ± 0.00250.1767 ± 0.00010.1780 ± 0.00110.1763 ± 0.0014

Table 3. Training cost on METR-LA.

Wall-clock training time on the same machine, mean over three seeds. GRCD is trained to convergence with validation-based early stopping, which stops at 6,250 updates on average. The baselines are reported at a fixed 5,000-update budget and at an extended 40,000-update budget. The 5,000-update arm uses fewer updates than GRCD but more wall-clock time, because consistency training and MeanFlow cost more per update.

MethodUpdatess / updateTotal trainingRelative to GRCD
GRCD (converged, early-stopped)6,2500.0874545.9 s1.00×
CT, 5,000-update budget5,0000.1198599.2 s1.10×
CT, 40,000-update budget40,0000.11874,746.2 s8.7×
MeanFlow, 5,000-update budget5,0000.1723861.3 s1.58×
MeanFlow, 40,000-update budget40,0000.17406,959.7 s12.7×

All 3 tables are in the Appendix. We thank the reviewer again for the suggestions.

Reviewer_7NH9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a a new insight in implementing generateive diffusion models for graph signals. Specifically, given that most graph signals have nonuniform spectral energy across graph Fourier modes (frequencies), instead of isotropic corruption that adds the same noise variance to each frequency, the authors propose a mode-dependent clock that gives every frequency tehe same conditional SNR.

Critical strengths
  1. The paper is theoreically sound in its analysis and algorithm development. The articulation of the key insight--that a mode-dependent clock is important for graph signal diffusion with nonuniform spectral energy distribution--is well described.
Critical weaknesses
  1. While the paper is technically sound, the proposal of a frequency-dependent clock appears to be a relatively minor contribution in the topic of graph signal diffusion. Without the proposal, graph diffusion still works just fine--it's just a little less efficient.

  2. The paper is written in a dense mathematical way that is not easily accessible to a non-epxert reader. Understanding the paper requires knowledge of many references cited throughout the manuscript, rather than embedding those knowledge pieces within the main content to facilitate comprehension of the main thesis.

Questions for authors
  1. While the reviewer agrees that most graph signals have non-uniform spectral energy across frequecies--especially when the graph is learned from signal observation via existing graph learning algorithms such as GLASSO and CLIME--what about the case when the distribution is strictly known, such as ideal low-pass or a decaying distribution such as f(λ)=(1+μλ)−1f(\lambda) = (1 + \mu \lambda)^{-1}? Most graph learning algorithms would result in distributions such as these on average.
1 comment

We are especially grateful for the reviewer's encouraging assessment of our work and the thoughtful suggestions for further improvement.

Reviewer: While the reviewer agrees that most graph signals have non-uniform spectral energy across frequencies--especially when the graph is learned from signal observation via existing graph learning algorithms such as GLASSO and CLIME--what about the case when the distribution is strictly known, such as ideal low-pass or a decaying distribution such as… ? Most graph learning algorithms would result in distributions such as these on average.

Thank you for this question. GRCD does not require the spectral profile to be estimated. Its forward process depends on the graph Laplacian eigenvalues and the noise schedule. The spectral profile enters through the Gaussian reference, which supplies the reverse-sampling initialization, the residual-score parameterization, and the exact modewise propagator, and can be specified analytically when known.

On a graph learned by GLASSO or CLIME, GRCD uses the learned Laplacian as given and fits one variance per mode in its eigenbasis, exactly as in our main experiments. A closed-form profile can replace the fit whenever one is known to be correct.

Our SBM datasets are generated from an analytically known decaying profile, v0(νi)=0.2+0.81+cνiv_0(\nu_i) = 0.2 + \frac{0.8}{1 + c\nu_i}, with c∈{1,4,16}c \in \{1, 4, 16\}, where νi\nu_i are the eigenvalues of the generator's combinatorial Laplacian, scaled by λmax⁡\lambda_{\max} and shifted, as defined in the appendix. In the original experiments, GRCD estimates this profile using Ledoit–Wolf, as on the real datasets. For this rebuttal, we repeated these experiments with the exact profile in place of the fit, using κ=2\kappa = 2, converged checkpoints, the exponential-residual solver on the ρ=3\rho = 3 grid used in the main tables, and three seeds. At NFE 4, 8, 16 and 64, the exact profile is better in all twelve comparisons, but by only 1.6% to 7.6% in aMMD, and every paired 95%95\% confidence interval includes zero. Hence the Ledoit–Wolf estimate is already accurate enough that knowing the profile exactly is worth at most about 7.6%.

We also tested an ideal low-pass setting for this response. It uses the same two-block SBM generator as our other synthetic settings (N=32N = 32), replacing the decaying profile with v0=1v_0 = 1 on the eight lowest graph frequencies and 10−210^{-2} above them, a small floor that keeps the data covariance non-singular. At NFE 4 on the ρ=3\rho = 3 grid, GRCD reaches aMMD 0.0429±0.00580.0429 \pm 0.0058, against 2.7530±0.00182.7530 \pm 0.0018 for GAD with its author-faithful sampler and 2.6691±0.04542.6691 \pm 0.0454 for EDM, both of which we trained on this setting for this response.

These results show that GRCD can use either known or estimated spectral profiles, and in the ideal low-pass setting the reviewer raised, it keeps its low-budget advantage over the graph-diffusion and isotropic baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.