acceptodds
Withdrawn from ICLR 2027

TopoFlowForge: Progressive Native Mesh Generation via Cascaded Flow Matching

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

3D content generation technology has significantly advanced the work of designers, as well as the 3D printing and gaming industries. However, it remains difficultto produce lightweight, editable, and topologically clean artistic content that isdirectly production-ready. To achieve this, we present TopoFlowForge, an artistic mesh foundation model that generates production-ready meshes. Specifically,TopoFlowForge decomposes the mesh generation process into vertices generationand their connectivity prediction, i.e., edges. We formulate vertices generationas a two-stage coarse-to-fine process and incorporate several effective loss functions to further enhance its performance. In the connectivity prediction stage, wepropose a simple yet effective method for estimating the connectivity affinity between vertices and additionally predict per-vertex normals, which determines thecorrect orientation of faces. Besides, we construct a large-scale dataset combininghand-crafted 3D assets with public high-quality topology datasets. Based on this, a carefully designed data curation pipeline is employed to filter the raw dataset,retaining only high-quality topology data for model training. Our model is trainedon the combined dataset and tested on both out-of-distribution hand-crafted set of3D assets and public datasets. Under image conditioning, TopoFlowForge surpasses autoregressive methods in both generation quality and inference speed,achieving state-of-the-art results among open-source mesh topology generators. We will release all code and weights together with a portion of our test dataset.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Generation, Artist Mesh Generation, Flow Matching

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_UYTSOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TopoFlowForge, a cascaded flow-matching framework for image-conditioned artistic mesh generation. The first stage generates a coarse occupancy grid representing vertex-support locations. The second stage progressively refines this support from 64364^3 to 5123512^3 using resolution-specific experts, producing the final vertex set. The third stage generates per-vertex topology latents and decodes them into vertex connectivity and normals, from which triangular faces and their orientations are recovered. Experiments on Toys4K and TE-388 show improvements over LATO.2 and EdgeRunner across the reported geometric, semantic-alignment, and rendering-based metrics.

Critical strengths
  1. Generating coarse vertex support before higher-resolution refinement is a reasonable design that separates global structure modeling from local detail generation.

  2. Introducing connectivity-informed occupancy losses in the refinement stage is well motivated. Table 2 provides evidence that these additional losses, considered together, improve the reported generation metrics.

  3. Jointly predicting vertex normals and connectivity provides an intuitive mechanism for determining face orientation during mesh reconstruction.

Critical weaknesses
  1. The benefit of normal prediction is not directly evaluated. Despite explicit orientation prediction being a highlighted distinction from LATO.2, the paper does not report normal consistency, face-orientation errors, or a comparison against conventional orientation-correction post-processing. These experiments are necessary to establish the practical benefit of the normal head.

  2. The reconstruction fidelity of the VAE components is insufficiently characterized. The paper should separately evaluate coarse-occupancy reconstruction and topology reconstruction, rather than relying only on end-to-end generation results. Comparisons with relevant representations from EdgeRunner, MeshGPT, MeshFlow, LATO, and LATO.2 would help clarify representation quality, where compatible evaluation protocols are available. Since the topology VAE keeps vertex coordinates fixed, its reconstruction should be assessed primarily through connectivity and face recovery rather than vertex-position accuracy alone.

  3. The novelty relative to closely related methods needs a more systematic explanation. The vertex-generation pipeline shares the broad coarse-structure and factorized-generation direction of LATO/LATO.2, while its progressive subdivision is closely related to hierarchical octree generation in Nexus. The authors identify some differences, particularly explicit normal prediction, but do not clearly separate inherited design choices from substantive new contributions. The benefit of using independent resolution-specific experts is also not established through comparison with a shared, resolution-conditioned model.

  4. The qualitative results do not fully substantiate the claimed stability and topology quality. Figure 5 shows competitive geometry, but the polygonal layouts remain less regular than those of some commercial baselines, as the authors also acknowledge. These examples alone do not establish that the remaining issues originate in vertex generation. Stage-wise diagnostics and direct topology metrics would help identify whether failures arise from vertex placement, connectivity prediction, or face recovery.

  5. The topology VAE lacks targeted ablations. In particular, the paper does not independently evaluate the proposed dual-projection connectivity head or joint normal prediction. A comparison against a shared-projection head, together with edge-prediction accuracy and orientation metrics, would help establish whether these architectural changes offer meaningful advantages over existing topology-generation designs.

Questions for authors

See weaknesses.

Reviewer_oUj8OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a progressive pipeline for native mesh generation. It first generates a sparse structure in a manner similar to TRELLIS, then generates mesh vertices conditioned on this structure, and finally constructs the topology by predicting edge connectivity. Several seemingly practical techniques are introduced during vertex and connectivity generation.

Critical strengths

The authors introduce several improvements to the native mesh generation pipeline, including progressive upsampling, boundary-aware weighted binary cross-entropy, and edge-normal prediction. The effectiveness of these techniques is demonstrated through ablation studies.

Critical weaknesses
  1. The paper contains basic proofreading issues. I list at least five examples below:
  • Page 5, L235: “As shwon in”
  • Page 7, L346: “trining dataset”
  • Page 7, L346: “network paramenters”
  • Page 8, L413: “commericial models”
  • Page 9, Fig. 6: “w/o Late-Stage Refinetment”
  1. The paper's approach to native mesh generation largely follows existing methods: first generating vertices and then predicting topological connectivity. I therefore consider its novelty limited and do not believe it would have sufficient impact on the community.

  2. The authors should be cautious about claiming that the generated meshes are “product-ready.” In my view, this claim requires validation using additional metrics for mesh geometric quality, including assessments of non-manifold edges and holes. Moreover, I consider the visual quality of edge layouts another criterion for being “product-ready.” However, the edge layouts shown in the qualitative results—for example, those of the chair and human in Fig. 4 are not visually pleasing.

Questions for authors
  1. I do not see how Stage 1 of this paper differs from Stage 1 of TRELLIS. My understanding is that TRELLIS.2 directly uses the Stage 1 checkpoint from TRELLIS.1 to predict sparse structures from images. Does this paper also directly use the TRELLIS.1 checkpoint for Stage 1?

  2. What is the fundamental difference between your proposed topology representation (Eq. 8) and the space-time representation used in existing methods, such as MeshFlow? What advantages does your representation offer over these alternatives?

  3. What advantages does your progressive upsampling approach to vertex generation offer over the octree-based approach used in Nexus? Although Nexus has not been open-sourced, I would appreciate the authors' insights behind their method design.

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

This paper proposes TopoFlowForge, an image-conditioned native mesh generation framework based on cascaded flow matching. The method decomposes mesh generation into three stages: coarse vertex generation, progressive vertex refinement, and topology reconstruction. It introduces topology-aware losses to improve vertex generation and jointly predicts vertex connectivity and normals to recover mesh faces with proper orientations. In addition, the authors construct a large-scale curated mesh dataset for training. Experiments on Toys4K and the proposed TE-388 benchmark demonstrate promising geometric accuracy, image alignment, and qualitative mesh generation results compared with existing methods.

Critical strengths

S1. Practical and Meaningful Research Problem.

The paper addresses the important challenge of generating native meshes with clean topology, which is valuable for practical 3D content creation and downstream applications such as editing, rigging, and animation.

S2. Promising Generation Quality.

The proposed method achieves strong quantitative results on Toys4K and TE-388, while the qualitative comparisons demonstrate good geometric details and overall shape completeness across different object categories.

S3. Effective Progressive Generation Framework.

The coarse-to-fine vertex generation strategy, combined with topology-aware supervision and joint edge-normal prediction, provides a reasonable framework for improving mesh generation quality.

Critical weaknesses

W1. Limited Comparisons with Recent Autoregressive Mesh Generation Methods.

The comparison with existing methods is relatively limited, particularly for autoregressive mesh generation. EdgeRunner is the only autoregressive baseline, despite being trained with relatively limited resources and data compared to more recent approaches. Stronger and more recent autoregressive methods, such as BPT and DeepMesh, are missing. Including these methods would provide a more comprehensive and convincing evaluation of the proposed framework.

W2. Limited Methodological Novelty.

While the proposed framework achieves promising results, its overall architecture appears to be largely built upon existing approaches, particularly LATO.2 and Nexus. The separation of vertex generation and topology prediction, together with the use of diffusion or flow-based models, has already been explored in these works. Although TopoFlowForge introduces joint edge-normal prediction, the methodological novelty beyond integrating and extending existing techniques remains somewhat limited.

W3. Insufficient Evaluation of Normal Prediction.

Although joint edge-normal prediction is one of the main contributions, the paper does not report Normal Consistency or other direct metrics evaluating predicted face orientations. Since accurate normal is central goals of the proposed method, additional evaluations of normal consistency and topological validity would better support the claimed advantages.

Questions for authors

Q1. Could the authors provide comparisons with more recent autoregressive mesh generation methods, such as BPT and DeepMesh? Since several recent autoregressive approaches support point-cloud-conditioned generation, it may be feasible to train or adapt these baselines under the same point-cloud conditioning setup to enable a more comprehensive comparison.

Q2. Given the architectural similarities between TopoFlowForge and existing methods such as LATO.2 and Nexus, could the authors provide a more controlled comparison by training their vertex generation modules with comparable model sizes and training settings? This would help clarify the advantages of the proposed progressive vertex generation strategy beyond differences in model capacity.

Q3. Could the authors include Normal Consistency (NC) or other quantitative metrics to directly evaluate the accuracy and consistency of the predicted surface normals? This would provide stronger evidence for the effectiveness of the proposed joint edge-normal prediction module.

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

Related papers

Discussion (0)

Sign in to comment.