acceptodds
Under review as a conference paper at ICLR 2027

Natural Graph Diffusion Solver

Abstract

Existing graph diffusion samplers operate under uniform time discretization, which does not reflect the non-uniform nature of distributional evolution in graph generative processes. In this paper, we establish a fundamental principle for graph diffusion generation: a natural sampler should progress at a constant rate of distributional evolution. We show that this principle admits a geometric characterization, revealing that equal distributional change is equivalent to traversing equal Fisher-Rao arc-length increments along the reverse diffusion trajectory. This connection bridges diffusion sampling and information geometry, providing a principled foundation for information-geometry-aware generation. Inspired by this insight, we introduce the Natural Graph Diffusion Solver (NGD-solver) that enforces constant distributional evolution along the reverse diffusion trajectory. NGD-solver derives an information-geometry-aware correction that can be seamlessly incorporated into the reverse sampling process. By adaptively calibrating reverse transitions according to the local geometric sensitivity of distributions, NGD-solver promotes a constant-rate and stable evolution in distribution space throughout the diffusion. NGD-solver can be seamlessly integrated into pretrained graph diffusion models at inference time. Extensive experiments on graph generation benchmarks demonstrate that NGD-solver consistently and significantly improves structural fidelity and generation quality.

Primary area:
generative models
Keywords:
Graph Diffusion

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_BHAHOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes the Natural Graph Diffusion Solver (NGD-solver), an inference-time modification for pretrained graph diffusion models. The method starts from the observation that uniform reverse-time steps need not produce uniform changes in the induced transition distributions. For continuous SDE samplers, it uses a local Fisher-Rao metric for Gaussian transitions to set the magnitude of each drift increment while retaining the direction suggested by the model.

Critical strengths

1.The paper focuses on a practical part of generative modeling: the sampling procedure can limit output quality even when the pretrained model is fixed. NGD-solver targets this stage directly and does not require retraining or changing the model architecture. 2.The proposed correction has a clear geometric motivation. The derivation links drift changes in a Gaussian reverse transition to a local Fisher-Rao length, and the implementation is simple enough to integrate into existing samplers. 3.The experimental results demonstrate the effectiveness of the proposed method.

Critical weaknesses

1.The proposed diffusion methods are limited. These more representative methods of discrete diffusion, such as DiGress, cannot be used. This significantly limits the application value of the proposed method. 2.The direction of each corrected drift is taken from the difference between the current model prediction and the previously applied drift, then normalized. The paper should discuss edge cases where this difference is zero or very small and whether normalization becomes unstable. 3.The methodological novelty of NGD-solver appears limited relative to prior information-geometric sampling work. [1] already applies the Fisher–Rao metric to graph diffusion and enforces approximately constant information-geometric progress along the sampling trajectory. NGD-solver’s main distinction appears to be how this constraint is implemented: it rescales drift magnitudes on a fixed time grid, whereas [1] adapts the time discretization. The connection to [2] further suggests that Fisher–Rao geometry has already been used to shape diffusion trajectories outside graph generation. The paper should more clearly establish what theoretical or practical advantage the fixed-grid drift correction offers over these alternatives, and compare directly with [1] under matched models, sampling budgets, and tuning protocols. [1] Information-Geometric Adaptive Sampling for Graph Diffusion. [2] Geodesic Diffusion Models for Efficient Medical Image Enhancement.

4.Following Weakness 3, there is no relevant baseline for comparison. And all the author's analyses before Eq.8 have nothing to do with graph diffusion and can be applied to any SDE field. Therefore, theoretically, the accelerated solver method in the image domain can also be applied for comparison.

Questions for authors

See Weakness

Reviewer_e7gUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Graph diffusion sampling can produce uneven distributional changes. This paper proposes NGD-solver, which rescales reverse-time drifts using a Fisher–Rao arc-length constraint without retraining. Experiments with continuous and discrete graph generators on general and molecular graph benchmarks report improved generation quality and low sampling overhead.

Critical strengths
  1. The Fisher–Rao framework offers an interpretable way to regulate distributional changes.
  2. Experiments cover multiple graph datasets, generative models, and numerical solvers.
  3. The method improves generation quality with little runtime overhead and can achieve competitive results with fewer sampling steps.
Critical weaknesses

1.The paper does not directly show that actual distributional changes become more uniform during sampling or that this explains the quality gains. 2. Comparisons with drift normalization, clipping, or simpler rescaling methods are needed to establish the benefit of Fisher–Rao geometry. 3. Some metrics become worse, while multi-seed significance tests are only reported for QM9 with GDSS.

Questions for authors

1.Does the modified sampler preserve the original diffusion process? Rescaling the drift changes the reverse dynamics. Is there any convergence or distribution-preservation guarantee? 2. Is Fisher–Rao geometry actually necessary? Would simple drift normalization or clipping achieve similar improvements under the same evaluation settings?

Reviewer_5CtFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses a problem in graph generation: taking equally sized time steps does not necessarily yield equal changes in probability distributions. The authors therefore propose adjusting each generation step to make distributional changes more uniform, as measured using Fisher–Rao geometry, while accounting for the noise level. The method does not require retraining the network, as the correction is applied only during generation.

Critical strengths
  • The paper clearly presents its motivation and accompanies the derivation with a description of the algorithm.
  • The work offers an interesting perspective on sampling: equal time intervals do not imply statistically equivalent changes.
  • The ability to apply NGD to pretrained models is particularly useful, both because it reduces the cost of adoption and because it makes experimental comparisons easier to interpret.
  • The experiments show the method’s effectiveness across different model families and domains, supporting its applicability beyond a single configuration.
Critical weaknesses
  • The constant C is selected separately for each dataset, but the selection procedure is unclear. Thus, it is difficult to assess how robust the benefits are without a good choice of C.
  • The method makes a local measure of distance between drift-parameterised transition distributions uniform. However, it remains unclear why this local control should improve the final distribution of the generated graphs. Further theoretical arguments or evidence would help clarify this connection.
Questions for authors
  • I may have missed this, but how is the parameter C selected?
  • Do the improvements persist when a single value of C is fixed for each model and used across different datasets?
  • I would suggest reporting standard deviations across multiple independent runs to help assess the stability of the results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.