acceptodds
Under review as a conference paper at ICLR 2027

DeltaFlow: Single-Stage Mesh Generation using Flow Matching with Laplacian Positional Encoding

Abstract

Triangle meshes remain the standard 3D representation for graphics and rendering. However, current explicit triangle mesh generation methods either rely on slow autoregressive methods that generate one triangle at a time or require a multi-stage pipeline with diffusion models that hinders optimization for downstream objectives. To fill this gap, we introduce DeltaFlow, a single-stage flow matching framework for unordered triangle soups using a Diffusion Transformer (DiT) architecture. Naive triangle soup generation uses unordered tokens to represent the generated triangles, leading to meshes with duplicated or missing triangles and poor connectivity. Instead, DeltaFlow introduces Laplacian positional encoding that encodes a mesh's dual-graph Laplacian coordinates onto a shared template for rotary position embeddings (RoPE), injecting a connectivity prior into the denoising process. Experiments show that DeltaFlow generates meshes of similar quality to those generated by other mesh generative methods and substantially tighter meshes than those generated by other flow-based methods. Furthermore, our native triangle flow matching model learns a generalizable prior that enables various downstream geometry processing applications, such as mesh inpainting and point-cloud-conditioned mesh generation.

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

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_3uoeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method called DeltaFlow, a single-stage flow-matching model built on a diffusion transformer (DIT) for direct triangle-mesh generation. The main contribution of this paper is a Laplacian positional encoding. The dual-graph Laplacian of the mesh is embedded onto a shared honeycomb template, and give each face a connectivity prior during denoising. This can makes meshes more regular and better connected than prior flow-based methods, with fewer duplicate faces. The learned priors also transfers to downstream tasks such as point-cloud-conditioned generation and mesh inpainting.

Critical strengths

The core idea is injecting connectivity through Laplacian positional encoding, which I think is sound and well-motivated. Mapping the dual-graph Laplacian onto a shared honeycomb template gives the model a connectivity prior that directly targets the main failure of triangle-soup generation, also duplicate and disconnected faces.

DeltaFlow uses a single-stage flow matching, which sidesteps face by face generation, while still keeping a positional prior. Comparisons show it produces better-connected meshes with fewer duplicate faces.

Critical weaknesses

The geometric-fidelity evaluation is incomplete. The paper mainly reports 1-NNA, a custom tightness score, and duplicate-face rate, but no overall Chamfer distance. The demand for CD applies specifically to the conditional settings for point-cloud-conditioned generation and mesh completion where ground-truth meshes do exist. Chamfer distance should be reported for those.

This paper is trained only on four categories in ShapeNet, and all qualitative results shown are shapes with simple mesh structures and geometries. However, most current mesh generation work is trained on more complex datasets. The data complexity demonstrated in this paper is far lower than that of current SOTA (MeshAnything, BPT, EdgeRunner) methods. This makes it impossible for me to judge the upper bound of the method's capability.

The reported MeshFlow figures appear not consistent with the original paper. Table 1 lists MeshFlow at 586M parameters and 2.8s inference, but MeshFlow's original paper report roughly 124M parameters and 0.877s? This is not a cosmetic mismatch, since the the paper claims to be smaller and faster than MeshFlow.

It seems the tightness metric lacks a ground-truth reference. Without this reference it is hard to judge how tight is good enough, or whether the reported improvements are meaningful in absolute terms.

Questions for authors

Could you please report Chamfer distance for the conditional tasks?

Could you add a No PE + No vertex weighting ablation to isolate the independent contribution of the Laplacian positional encoding?

Please expand the dataset by incorporating more diverse shapes to demonstrate the upper-bound performance of the method and compare it against SOTA approaches.

2 comments

Dear Reviewer 3uoe,

With one day left in the discussion period, we would greatly appreciate it if you could take a look at our responses and the additional results we provided to address your concerns. Please let us know whether they resolve them, or if anything remains that we can clarify in the time left. If so, we would be grateful if you would consider raising your score.

Thank you again for your time.

Reviewer_3uoeOpenReview ↗

Concerns About the More Diverse Datasets

Thank you very much for your reply. The response addressed most of my questions except one. I still have major concerns regarding your claim in Fig. 12 about more diverse datasets. It can be clearly seen that the shapes listed in the Figure have simplest geometries in Objaverse. Most of them are box-like shapes with very few faces. This cannot demonstrate the method's ability to generalize to complex shapes.

The field of mesh generation has developed rapidly in recent years, with increasingly higher face counts and complexity. I think that this kind of extremely sparse and simplistic mesh generation is far from the current development trends.

Also, I do not think that the text and point cloud conditions can serve as a reason for insufficient complexity. These are merely differences in input and does not affect the representational capacity of the network. Overall, I believe the proposed method is indeed interesting and innovative and very insightful, but the extremely simple results on a few categories really run counter to current development trends in the mesh generation field.

Reviewer_1qyiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces DeltaFlow, a mesh generation method that performs flow matching directly on triangle coordinates to mitigate gaps, duplicate faces, and loose connectivity in generated triangle sets. The method first computes two-dimensional Laplacian spectral coordinates from the face-dual graphs of training meshes and matches these coordinates to a shared honeycomb template, assigning each triangular face a fixed positional identity. The network incorporates template positions into face attention through RoPE and combines vertex-slot weights with global vertex attention to predict vertex velocities. At generation time, only the face count and the corresponding template are required, without access to the target mesh’s adjacency information. Experiments on four ShapeNet categories demonstrate improvements in mesh tightness and further illustrate applications to mesh completion and point-cloud-conditioned generation without retraining.

Critical strengths
  1. The research problem is clearly defined, and the method directly addresses it. The paper addresses inadequate local connectivity in direct triangle generation and proposes coordinating triangle motion through stable face identities and neighborhood priors. Modeling triangle coordinates directly also allows constraints on observed faces and geometric losses to act directly on the generated outputs, giving the approach practical value.
  2. The shared template design addresses the discrepancy in information availability between training and inference. The spectral coordinates of ground-truth meshes are used only to assign faces to slots in the training data, while the network uses template coordinates shared by meshes with the same face count. Consequently, inference does not depend on the connectivity of the unknown target mesh. This design provides a reasonable way to translate the topological information of training meshes into an inductive bias for the generative model.
  3. The main experiments show relatively consistent improvements in tightness, supported by component ablations. In Table 1, DeltaFlow substantially reduces tightness error relative to MeshFlow across the four categories, while largely preserving the overall quality of the generated distribution. Table 2 further shows that removing positional encoding or vertex attention degrades performance, supporting the practical contribution of these components to the current system.
Critical weaknesses
  1. The paper does not establish the specific contribution of the Laplacian template design or its applicability to complex topologies. The main quantitative ablation compares the full method with a variant without positional encoding. This demonstrates the usefulness of positional encoding but does not distinguish whether the gains arise from generic fixed face identities or from the neighborhood information provided by Laplacian spectral encoding. Moreover, compressing mesh neighborhoods into two-dimensional spectral coordinates and matching them to a fixed template may distort neighborhood relationships. Although the paper acknowledges this limitation, it neither quantifies the distortion nor analyzes its relationship to topological complexity, the validity of generated meshes, or generation diversity.
  2. The paper does not establish scalability to meshes with high face counts. The experiments do not systematically report how generation quality changes as the face count increases, nor do they adequately validate connectivity quality on complex meshes. Face-count conditioning and interpolation between face-count embeddings show that the model can accommodate different face counts, but they do not establish that it maintains performance at higher face counts or beyond the training range.
  3. The contribution of post-processing to connectivity metrics is insufficiently clear. The appendix states that the outputs undergo coordinate quantization, threshold-based merging, and removal of duplicate and degenerate faces. These operations may directly affect tightness. However, the stage at which the metrics are computed, the consistency of thresholds across methods, and the magnitude of improvement from raw outputs to final meshes remain insufficiently specified. Furthermore, the metrics discussed in the paper do not fully capture issues such as non-manifold geometry, self-intersections, or incorrect connectivity. The current results therefore offer only a limited assessment of the topological quality of the final meshes.
Questions for authors
  1. Could the authors provide a quantitative comparison between Laplacian and lexicographic face assignment, extending the qualitative comparison in Figure 4? Using the same template and RoPE, with matched architecture, training budget, sampling, and post-processing, would help isolate the benefit of the spectral assignment from that of fixed face identities.
  2. Could the authors quantify neighborhood preservation after spectral embedding and template matching on the existing evaluation set, and examine its relationship to tightness? If the data permit, a breakdown by connected-component count or topological complexity would help clarify when the template provides a useful neighborhood prior.
  3. What face-count ranges are used for training and evaluation? Could the authors report generation quality, runtime, and peak GPU memory at several face counts within the evaluated range? Please also clarify whether inference beyond the training range is supported or remains untested.
  4. The paper states that a released MeshFlow checkpoint is used, but Table 1 lists 586M parameters, whereas the public MeshFlow paper reports 124M. Could the authors identify the exact checkpoint and explain the parameter-counting convention, including whether auxiliary networks are counted? Please also clarify whether the checkpoint was modified or fine-tuned, whether the learned denoiser was used, and the sampling configuration.
  5. Appendix A.2 specifies that duplicate and degenerate faces are removed before computing tightness. Is tightness evaluated before or after coordinate quantization and vertex merging, and are the merging thresholds consistent across methods? Could the authors provide a comparison before and after these operations while keeping the metric’s filtering rules consistent? An additional diagnostic for non-manifold edges or self-intersections would help assess quality beyond tightness.
1 comment

Thank you for your response; I will raise my score. However, I would like the authors to discuss more recent diffusion-based work, including MeshFlow (CVPR 2026), TriFlow, MeshBDF, LATO, and LATO.2.

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

This paper introduces DeltaFlow, a single-stage flow-matching mesh generation pipeline. The central contribution of the paper is to introduce Laplacian spectral embeddings onto a the network's latent space. This provides a connectivity prior that reduces gaps and improves triangle alignment. Its a simple but powerful contribution, resulting in a method that achieves improved generation quality against prior work.

Critical strengths
  • The paper has a simple and elegant central idea. It maps Laplacian spectral embeddings onto a shared network space, introducing a connectivity prior that remains available during generation, without requiring the target mesh topology.
  • The method presents strong results, beating previous baselines
  • I really liked the way the authors materialized their spectral encoding through the mapping of mesh-specific Laplacian spectral coordinates onto a shared honeycomb template. This transforms the connectivity-informed neighborhood structure into a positional encoding that is available even when sampling during inference.
  • The method is efficient enough to support downstream applications that rely on posterior sampling optimization.
Critical weaknesses
  • The paper does not evaluate the face orientation and normal consistency. There are no metrics that establish consistent triangle winding or outward-facing normals. Including a discussion about this would strengthen the paper.
  • The vertex/faces representation before equation 9 needs clarification. As I understand, the triangle coordinates are encoded into vertex features. They are indexed by the number of faces F x 3, which this means that each face has distinct vertices, and a final mesh would have duplicated vertices across edges. While the method might actually benefit from such discretization, this should be stated clearer on the final manuscript.
  • The spectral contribution could be better ablated. Comparing against no positional encoding conflates distinguishing face identities with encoding connectivity. Therefore quantitative comparisons against lexicographic and random face-to-template assignments would strengthen the central claim.
Questions for authors
  • Could you show representative failure cases, particularly where mapping spectral coordinates onto the shared template poorly preserves face neighborhoods? Which mesh structures are most challenging? What are future directions that this method can be improved?

Comments on the paper

1 comment

Dear Reviewers,

The authors have submitted their rebuttal to address your concerns. As the discussion deadline approaches, please review their response and share any remaining concerns or questions soon so the authors have enough time to respond.

Thank you, -AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.