acceptodds
Under review as a conference paper at ICLR 2027

Geometric Flow Matching

Abstract

We propose a flow-based generative model for 3D surface generation that leverages geometric flows within the flow matching framework. Instead of relying on generic stochastic interpolation paths, we construct data-dependent trajectories using a curvature-driven geometric evolution that progressively smooths surfaces while preserving mesh topology. This formulation defines structured transport paths between simple base shapes and complex target geometries, enabling the model to learn geometry-aware transformations. Our approach operates directly on triangle meshes and produces coherent deformation sequences governed by surface curvature. We evaluate our method on 3D hand and face mesh datasets and compare it against alternative flow construction strategies, demonstrating that curvature-driven trajectories yield more realistic, geometrically faithful meshes.

Primary area:
generative models
Keywords:
Flow Matching, 3D shape Generation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_Ds7UOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Geometric Flow Matching for generating 3D triangle meshes. It uses conformal Willmore flow to smooth training meshes into simple base shapes and constructs deterministic trajectories for flow-matching supervision. A graph neural network learns to reverse these trajectories by predicting vertex movements. Experiments on hand and face meshes evaluate reconstruction accuracy and mesh quality.

Critical strengths

Strengths

  1. Meaningful motivation. The paper addresses an important challenge in direct mesh generation: vertex movements can produce inverted triangles, self-intersections, and folds. Geometry-aware trajectories offer a reasonable approach to improving mesh validity.
  2. Clear formulation and useful controlled comparisons. The integration of geometric flows into flow matching is presented clearly. Comparisons using the same network with different forward trajectories help isolate the effect of the proposed path design.
Critical weaknesses

Weaknesses

  1. Limited methodological novelty. The main contribution combines an existing geometric flow with an established flow-matching framework and a conventional graph neural network.
  2. Limited generality. Hand and face models are trained separately, with fixed vertex counts and connectivity. The geometric-flow setup also assumes closed surfaces with sphere-like topology. These restrictions narrow the method’s applicability, and its effectiveness on more diverse categories and mesh structures remains unclear.
  3. Limited comparison with recent 3D generative methods. The comparisons with prior generative models largely focus on older methods and do not adequately position the proposed approach against recent 3d generations. Comparisons that account for differences in conditioning and mesh representation would help establish whether the method offers meaningful advantages in geometric quality, mesh validity, or efficiency. Furthermore, the demonstrations are limited to hands and faces, leaving its effectiveness on more diverse and complex shapes unclear.
Questions for authors

As weakness.

2 comments

We thank the reviewer for their evaluation and feedback. We address the three concerns below.

W1. Methodological novelty

Our main contribution is not a new geometric-flow solver or neural architecture, but a geometry-driven approach to constructing transport paths for flow matching on meshes. Rather than perturbing vertices independently, we use geometric evolution to define coherent intermediate shapes and corresponding base states. This allows the learning process to follow geometrically meaningful deformations rather than arbitrary vertex movements. Importantly, the framework is not restricted to conformal Willmore flow. Willmore flow provides one example of a geometric evolution that gradually simplifies complex surfaces toward a near-spherical state. Other suitable geometric flows could also be incorporated into the same framework, allowing transport trajectories to be designed using geometric principles rather than shape-specific deformation rules. We tested this directly: constructing training trajectories via straight-line interpolation between source and target produced severely self-intersecting intermediate meshes, before training was even possible. This observation confirms that geometric structure in the transport path, not just a deterministic source-target correspondence, is necessary.

W2. Generality and mesh topology

We acknowledge that our current implementation assumes fixed connectivity and vertex count within each dataset, and that the geometric-flow construction is applied to closed, sphere-topology meshes. The hand and face models are trained separately, and we do not claim that the trained models generalize directly across different mesh structures or topologies. However, the underlying geometric principle is not specific to hands or faces. Both datasets use curvature-driven evolution without requiring different category-specific deformation strategies. Extending the framework to other mesh categories, more complex topologies, and variable connectivity remains an important direction for future work.

W3. Comparisons with recent generative methods

Our experiments are designed to study the effect of transport construction by keeping the network architecture and training procedure fixed across baselines, using mesh samples drawn from the same underlying distribution. This allows us to focus on differences in transport construction without introducing differences in model capacity. We agree that comparisons with recent 3D generative methods under a shared evaluation protocol would provide a broader assessment. However, differences in mesh representations, conditioning, and generation tasks make direct numerical comparisons difficult. Accordingly, Table 4 is intended as contextual information rather than evidence of superiority over those methods.

The current implementation relies on fixed connectivity and vertex count, restricting the method's generality. Furthermore, it does not demonstrate clear superiority over current direct 3D mesh generation methods. For these reasons, I maintain my score.

Statement: I wrote this review myself.

Reviewer_QFkdOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Geometric Flow Matching (GFM), which constructs mesh deformation trajectories using conformal Willmore flow and trains a graph neural network to reverse the smoothing process. Experiments evaluate reconstruction on fixed-connectivity hand and face meshes, with additional generation experiments using PCA-based sampling and interpolation of smoothed base states.

Critical strengths
  1. The geometric motivation is clear: curvature-driven trajectories offer a meaningful alternative to independent vertex perturbations.
  2. Results on face meshes show improvements over the proposed baselines, and the qualitative reconstructions demonstrate smooth deformation trajectories.
  3. The authors transparently report weaker hand reconstruction results and provide the complete sampling gallery, including failures.
Critical weaknesses
  1. The main experiments reconstruct targets from their own smoothed meshes, which may retain target-specific information. The PCA and interpolation experiments lack sufficient quantitative evaluation of realism, diversity, and coverage.
  2. The controlled experiments include only two author-constructed baselines, while Table 4 reports older methods under incompatible evaluation settings. GFM also underperforms Sphere-Sweep on all four hand reconstruction metrics, leaving its empirical advantages insufficiently established.
  3. Fixed connectivity does not guarantee intersection-free or nondegenerate meshes, and the learned reverse process has no stated mechanism enforcing these properties. Moreover, deviation from ground-truth self-intersection rates can obscure geometric defects when the ground truth itself contains intersections.
  4. The method assumes a fixed vertex count and connectivity, restricting generation to deformations of a prescribed mesh topology. This prevents valid generation across different genera or numbers of connected components. Furthermore, the experiments on relatively low-resolution hand and face meshes do not establish whether the approach scales to substantially more detailed and geometrically complex shapes.
Questions for authors
  1. Can the authors provide quantitative evaluations of independently generated meshes, including failed samples, and numerical evidence supporting the nearest-neighbor claims?
  2. Which geometric properties are guaranteed, and can the authors report absolute self-intersection rates and the proportion of completely intersection-free outputs?
  3. Beyond Figures 2–3, could the authors provide representative successful and failed reconstructions for GFM and the baselines, using the same target meshes and viewpoints with a consistent selection criterion? In particular, examples where Sphere-Sweep outperforms GFM on hand meshes would help clarify the geometric differences behind the quantitative results.
2 comments

We thank the reviewer for their detailed feedback and address their concerns below.

W1 and Q1. Independent generation and quantitative evaluation

Reconstruction from target-specific smoothed states does not, by itself, establish independent generative capability, since these states may retain target information.

To evaluate generation beyond reconstruction, we construct new base states through PCA sampling and interpolation of smoothed training face states, then apply the trained reverse model without retraining. Among 100 PCA samples and 25 interpolated samples, 82% and 96%, respectively, satisfy our geometric validity criteria.

We further evaluated the generated faces using Coverage (COV), Minimum Matching Distance (MMD), and 1-Nearest-Neighbor Accuracy (1-NNA), based on Chamfer Distance against 200 held-out real identities. For all 125 generated samples, including unsuccessful samples, we obtain:

  • COV: 3.40×10−13.40 \times 10^{-1}
  • MMD: 8.61×10−38.61 \times 10^{-3}
  • 1-NNA: 8.99×10−18.99 \times 10^{-1}

Results show partial coverage of the held-out face distribution, while the high 1-NNA indicates that generated and real samples remain distinguishable. Our nearest-neighbor analysis also suggests that the generated faces are not simply reproducing training identities, as their distances to the training set are comparable to those of held-out real identities.

W2. Comparative evaluation and reconstruction performance

Our main goal is to investigate the effect of geometry-driven transport on mesh learning. We therefore use the same network architecture and training procedure across baselines, with mesh samples drawn from the same underlying distribution, to focus on the effect of transport construction.

GFM substantially outperforms the Gaussian-noise baseline on both hands and faces. On faces, it also achieves better reconstruction accuracy than Sphere-Sweep, with a Chamfer Distance of 1.36×10⁻³ compared with 2.25×10⁻³ for Sphere-Sweep.

Sphere-Sweep performs better on hand reconstruction, but it was specifically designed for hand meshes, using a centering mechanism based on PCA and identification of the palm and finger-like regions. In contrast, Willmore flow uses the same general geometric principle for both hands and faces, without requiring category-specific deformation rules.

Comparisons with established mesh-generation methods under the same protocol would provide a broader assessment. Table 4 is intended as a contextual comparison rather than a direct performance ranking, since the datasets and tasks differ.

W3 and Q2. Geometric validity and self-intersections

Fixed connectivity alone does not guarantee geometrically valid meshes. While the connectivity remains unchanged throughout the trajectory, the learned reverse process does not explicitly enforce intersection-free or nondegenerate geometry.

Our original evaluation reports deviations from ground-truth self-intersection rates, measuring how much reconstructions differ from their targets. We agree absolute rates give a clearer picture of geometric validity. For the same 125 samples, we additionally obtain:

  • Mean absolute self-intersection rate: 0.55%
  • Median absolute self-intersection rate: 0.17%
  • Completely intersection-free outputs: 44/125 (35.2%)

These results indicate that the generated meshes generally have low proportions of intersecting faces, but many are not completely intersection-free. Our geometric validity criteria also allow a small number of intersections, so satisfying these criteria does not imply an intersection-free mesh.

W4. Fixed topology and scalability

We acknowledge that our current implementation assumes fixed connectivity and vertex count within each dataset. Consequently, it generates geometric deformations of a prescribed mesh topology rather than shapes with arbitrary topology, different genera, or numbers of connected components.

Our experiments on hands and faces demonstrate the applicability of the same geometric transport principle to two different shape categories. However, we have not evaluated substantially higher-resolution or more geometrically complex meshes. Scaling to these settings, particularly considering the computational cost of offline Willmore-flow preprocessing, remains an important limitation of the current work.

Q3. Visual comparisons of reconstruction results

We would like to clarify that our held-out reconstruction experiments did not exhibit distinct successful and failed cases of the kind observed in independent PCA-based generation. All held-out test meshes were reconstructed, with differences in reconstruction accuracy across methods.

Sphere-Sweep achieves better average reconstruction accuracy on hands, as reported in our quantitative results. However, this does not imply that GFM fails on particular hand configurations. The test meshes vary in hand pose and spatial configuration while sharing the same underlying topology and anatomical structure.

The current implementation relies on the assumption of fixed connectivity and vertex count, estricting generation to deformations of a prescribed mesh topology. For these reasons, I maintain my score.

Reviewer_pq62OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Geometric Flow Matching (GFM) for fixed-connectivity triangle meshes. The method uses conformal Willmore flow to construct trajectories that progressively smooth target meshes into near-spherical configurations. A graph neural network learns reverse per-vertex displacements along these trajectories. Experiments on MANO hands and MediaPipe-derived face meshes compare this construction with Gaussian perturbations and Sphere-Sweep. Additional face experiments explore generation from PCA-sampled and interpolated base states.

Critical strengths
  1. Curvature-driven trajectories provide a reasonable geometric prior for learning mesh deformations.
  2. The method operates directly on explicit meshes, and the appendix transparently includes sampling failures.
Critical weaknesses
  1. The comparative evaluation is substantially insufficient. The main comparisons use the same network with Gaussian Noise, Sphere-Sweep, or the proposed geometric trajectories. These are primarily internal transport-path comparisons, rather than comparisons against established mesh generation methods. Table 4 reports results from different datasets and tasks and therefore cannot establish relative performance. The independent-sampling experiments also lack competing generative baselines. Consequently, the paper does not adequately demonstrate its value relative to existing approaches.
  2. The methodological novelty is limited. The approach combines an existing conformal Willmore-flow solver, graph convolutions, and a standard flow-matching regression objective. Its main contribution is the construction of geometric trajectories and deterministic source–target coupling. While reasonable, this incremental design requires stronger evidence of its specific benefit. In particular, comparing alternative paths between identical source–target pairs would help separate the effect of trajectory geometry from differences in initialization and retained target information.
  3. The resulting advantages remain unclear. Improvement over independent vertex noise supports the motivation, but does not establish competitiveness with existing generators. GFM also underperforms Sphere-Sweep on all four hand reconstruction metrics. Most quantitative experiments reconstruct targets from their own smoothed states, while independent generation is evaluated only on faces using 100 PCA samples and 25 interpolations. The limited validity metrics do not sufficiently establish generation quality or diversity. Given the offline trajectory construction and fixed-connectivity constraints, the paper should demonstrate a clear benefit that justifies these requirements.
Questions for authors
  1. Can the authors compare against representative mesh generators under a consistent dataset, task, and evaluation protocol?

  2. Can they isolate the trajectory’s contribution using identical source–target pairs, and identify the principal advantage supported by the results?

2 comments

We thank the reviewer for the careful reading and helpful comments. We would like to clarify the main contribution of GFM and explain some of our experimental choices. We address each of the reviewer's comments and questions below.

W1. Comparison with existing mesh-generation methods

Our main goal is to investigate whether geometry-driven transport can improve mesh learning compared with conventional noise-based approaches. For this reason, we use the same network architecture and training procedure across all baselines, with mesh samples drawn from the same underlying distribution. This allows us to study the effect of the transport construction without introducing differences in model capacity or training conditions. Our results show that GFM outperforms the Gaussian-noise baseline on both hands and faces, and achieves better reconstruction accuracy than Sphere-Sweep on faces. Importantly, these results are obtained on held-out test meshes that were not seen during training.

We agree that comparisons with established mesh-generation methods under the same dataset and evaluation protocol would provide a broader assessment of GFM. However, existing approaches often differ in their mesh representations, conditioning, and generation tasks, making direct comparisons challenging. Table 4 is therefore intended to provide context from previous work, rather than a direct performance ranking.

For independent generation, we additionally report Coverage, MMD, and 1-NNA against 200 held-out real face identities. While these metrics provide a broader evaluation of generative quality, we acknowledge that comparisons with established generators under the same protocol would be valuable.

W2. Methodological novelty and the role of geometric trajectories

We agree that the individual components, including the Willmore-flow solver, graph convolutions, and flow-matching regression, are established techniques. Our main contribution is their combination into a geometry-driven transport framework, where geometric evolution defines both the transport trajectory and the corresponding base state.

Unlike generic noise-based transport, our approach uses curvature-driven evolution to construct smooth, geometrically meaningful intermediate meshes. The framework is not limited to Willmore flow; other suitable geometric evolutions could also be used to construct such trajectories.

We also explored a simpler alternative using straight-line interpolation between the original mesh and a radially projected spherical mesh. However, the intermediate meshes developed severe self-intersections before training, motivating our use of curvature-driven trajectories.

W3. Reconstruction performance and independent generation

Our results show that GFM substantially outperforms the Gaussian-noise baseline on both datasets. On faces, it also achieves better reconstruction accuracy than Sphere-Sweep, with a Chamfer Distance of 1.36×10−31.36 \times 10^{-3}, Normal Consistency of 0.990, and a self-intersection rate of only 0.12% on held-out test meshes. For hands, Sphere-Sweep performs better across all four reconstruction metrics. However, Sphere-Sweep was specifically designed for hand meshes, with a centering mechanism based on identifying the palm region using PCA and separating finger-like regions. We applied the same method to faces without changing its underlying construction. In contrast, Willmore flow uses the same general geometric principle for both hands and faces, without requiring separate deformation rules for each category.

We recognize that reconstruction from target-specific smoothed states does not, by itself, establish independent generative capability. To investigate independent generation, we sampled new base states using PCA and interpolation and applied the trained reverse model without retraining. Among 100 PCA samples and 25 interpolated samples, 82% and 96%, respectively, satisfy our geometric validity criteria. Our nearest-neighbor analysis also shows that generated samples have distances to the training set comparable to those of held-out real identities, providing preliminary evidence against simply reproducing training examples.

To further evaluate independent generation, we computed Coverage (COV), MMD, and 1-NNA using Chamfer Distance against 200 held-out real face identities. For the combined set of 125 generated samples, including unsuccessful samples, we obtain:

  • COV: 3.40×10−13.40 \times 10^{-1}
  • MMD: 8.61×10−38.61 \times 10^{-3}
  • 1-NNA: 8.99×10−18.99 \times 10^{-1}

These results show partial coverage of the held-out face distribution, while the high 1-NNA indicates that the generated and real distributions remain distinguishable.

Thank you for your reply, but I still have concerns about the lack of method novelty, so I will keep my score.

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

The paper presents a well-motivated application of flow matching to 3D mesh generation. Its central contribution is to construct transport trajectories through curvature-driven geometric smoothing, incorporating surface structure directly into the learning process. This provides coherent supervision for recovering detailed geometry and makes the intermediate transformations interpretable. The experiments show substantial improvements over the Gaussian-noise baseline and favorable face reconstruction results compared with Sphere-Sweep, supported by evaluations of both reconstruction fidelity and mesh quality. Sampling and interpolation experiments also provide preliminary evidence that the learned reverse dynamics extend beyond precomputed reconstruction trajectories. Although broader benchmarking would strengthen the evaluation, the work offers a promising connection between classical geometry processing and flow-based generative modeling.

Critical strengths
  1. Well-motivated geometric design. The paper addresses the coupling between neighboring mesh vertices by constructing curvature-driven deformation trajectories. This provides a meaningful geometric prior for learning surface transformations and encourages coherent intermediate meshes.

  2. Simple and interpretable approach. The method connects established geometric smoothing techniques with a straightforward displacement-prediction objective. The intermediate surfaces make the learned transformation easy to inspect, providing a clear view of how detailed geometry is progressively recovered.

  3. Encouraging results and informative evaluation. GFM substantially improves over the Gaussian-noise baseline and achieves favorable reconstruction results on faces compared with Sphere-Sweep. The evaluation considers curvature, normal consistency, and self-intersections alongside reconstruction distances, providing a broader assessment of surface quality. Experiments with newly sampled base states also provide preliminary evidence of generalization beyond precomputed reconstruction trajectories.

Critical weaknesses
  1. The promising generative results could be strengthened through broader evaluation. The reconstruction experiments demonstrate that the model can recover detailed geometry from smoothed meshes, and the continuous-prior experiment provides encouraging evidence beyond target-specific reconstruction. Since the main experiments start from states that may retain target information, additional evaluation of independently generated samples would help establish the broader generative capabilities. Extending continuous-prior sampling to hands and assessing realism, diversity, and distribution coverage would be particularly informative, alongside the geometric validity measures.

  2. Additional controlled comparisons would help clarify where geometric flow provides the greatest benefit. Using a shared architecture across transport strategies is a useful experimental design, and the face results are encouraging. Sphere-Sweep achieves stronger reconstruction results on hands, suggesting that the relative benefits may depend on the shape category and information retained by the forward process. Comparisons with simpler smoothing methods, using similarly informative base states, would help identify the specific contribution of curvature-driven trajectories. Evaluating an established mesh-generation method under the same protocol would also complement Table 4’s contextual comparison.

  3. Clarifying the connection between the implemented updates and the flow-matching formulation would strengthen the method’s presentation. The time-independent displacement predictor provides a simple implementation, while the mathematical formulation describes a general time-dependent velocity field. Explaining how this parameterization relates to the continuous-time formulation, including the role of the timestep, would make the connection clearer. Experiments with different numbers of updates could further characterize the tradeoff between sampling cost and quality and demonstrate the practical flexibility of the learned dynamics.

Questions for authors

Section 3.5 states that the network receives neither time nor step index. What motivated this choice, and do your observations suggest that the current mesh geometry provides sufficient information to determine the appropriate next displacement throughout the trajectory?

1 comment

We thank the reviewer for the positive assessment of our model and the constructive suggestions for strengthening the evaluation and clarifying the methodology. We address each of the reviewer's comments and questions below.

W1. Generative Evaluation

We agree that geometric validity alone is not enough to evaluate generative quality, especially diversity and distribution coverage. Our PCA sampling and interpolation experiments show that GFM can generate faces from new base states rather than only reconstructing held-out test meshes. To further evaluate these samples, we computed Coverage (COV), Minimum Matching Distance (MMD), and 1-Nearest-Neighbor Accuracy (1-NNA) using Chamfer Distance against 200 held-out real face identities. For the combined set of 125 generated samples, including unsuccessful ones, we obtain:

  • COV: 3.40×10−13.40 \times 10^{-1}
  • MMD: 8.61×10−38.61 \times 10^{-3}
  • 1-NNA: 8.99×10−18.99 \times 10^{-1}

These results show partial coverage of the real face distribution, while the high 1-NNA indicates that the generated and real distributions remain distinguishable. We recognize that broader generative evaluation would help further establish the capabilities of GFM.

W2. Benefits of geometric trajectories

An important point is that our framework is not limited to Willmore flow. Willmore flow is one example of a geometric evolution that gradually simplifies complex surfaces toward a canonical shape, which in our case is a near-spherical mesh. Using suitable geometric flows has two main advantages. First, unlike heuristic methods such as Sphere-Sweep, which was specifically designed for hand meshes, geometric flows use properties such as curvature and local surface geometry without requiring shape-specific deformation rules. Second, they provide smooth, geometrically meaningful trajectories that help maintain mesh quality and reduce problems such as self-intersections and distorted faces. Our results with Willmore flow demonstrate the potential of this approach, but the framework could also be extended to other geometric flows that satisfy the required conditions. Regarding the stronger hand reconstruction results of Sphere-Sweep, its performance is consistent with its hand-specific design. In contrast, Willmore flow uses the same general geometric principle for both hands and faces, without requiring a different deformation strategy for each category. We also explored straight-line interpolation between the original mesh and a spherical mesh as a simpler alternative. However, the intermediate meshes developed severe self-intersections, since directly interpolating vertex coordinates does not account for the geometric relationships between neighboring vertices. This observation further motivated our use of curvature-driven evolution.

W3 and Question: Time-independent prediction

We thank the reviewer for this insightful question regarding our model's parameterization and its connection to the general flow-matching formulation.

Connection to flow matching. Our model learns a time-independent discrete update that predicts per-vertex displacements between consecutive mesh states, representing the velocity integrated over each step. ChebConv layers capture local geometric information to produce spatially varying predictions based on the current mesh geometry, without requiring an explicit time or step-index input.

Motivation and limitations. Our initial motivation was to investigate whether the current mesh geometry alone could provide enough information to predict the next displacement. We therefore applied the same time-independent predictor throughout the reverse trajectory.

Our reconstruction results suggest that this approach can learn useful reverse dynamics, but they do not establish that geometry alone always determines the appropriate displacement. We agree that including time or step index could help distinguish different stages of the evolution, reduce ambiguity, and potentially improve learning efficiency and reconstruction accuracy.

We have not performed a controlled comparison with time-conditioned models or systematically tested different numbers of inference steps. Therefore, we cannot yet assess their effects on performance or the tradeoff between computational cost and reconstruction quality. We consider these promising directions for future work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.