acceptodds
Withdrawn from ICLR 2027

LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

11 views

Abstract

Flow matching over carefully designed latent representations has recently emerged as a powerful paradigm for topology-aware mesh generation. Existing approaches, however, model vertices and connectivity jointly in a joint latent space, entangling continuous vertex geometry with discrete combinatorial structure; this complicates flow learning and manifests as drifting vertices and broken surfaces. We present LATO.2, a factorized flow matching framework that decomposes mesh generation into a vertex flow followed by a connectivity flow conditioned on the realized vertices, with both stages anchored to a shared coarse voxel scaffold. Dedicated VAEs underpin the two stages, recovering vertices at sub-voxel precision and embedding discrete connectivity into a continuous latent space. We demonstrate two advantages unique to this factorization: (i) part-wise generation, in which the scaffold is partitioned and each part synthesized at full latent capacity, yielding substantially higher-resolution meshes than a monolithic latent permits; and (ii) topology-adaptive editing, in which manipulating first-stage vertices induces the corresponding connectivity without re-optimization. Experiments show that LATO.2 surpasses state-of-the-art topology-aware mesh generators in geometric fidelity and connectivity quality.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3d shape representation, 3d generation, diffusion model, mesh topology

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_4hXtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces a two-step explicit mesh generation method. The first step (v-flow) generates vertices and the second step (t-flow) generates the connects in-between vertices.

The main benefits are high-quality, controllability.

The main concrens are a) unclear explanation why factorization is resonsible to the benefits. b) limited novelty of the core factorization idea.

Critical strengths

The overall improvements across all four metrics in Table 1.

Critical weaknesses
  1. Insufficent evidence for the benefits of factorization. Although the factorization is natural and the authors claim that seperating vertex and topology simplifies flow learning and reduce broken surfaces. However, the paper lacks controlled comparison between joint and factorized flow models. Since the proposed methods also introduce several architectual changes, it's unclear whether the observed improvments are directly from factorization.

  2. Unclear effectiveness of vertex-conditioned topology generation. The authors claim that realized vertices simplifies connectivity prediction. However, similar vertex can correspond to different valid topologies. For example, a empty bottle (has inner surface and outer surface) and a solid bottle (only outer surface), similar vertex structures but different topologies. It remains unclear how the proposed method handles such cases.

  3. Limited conceptual novelty of the core factoization idea. The decomposition of mesh generation into vertex-topology prediction has been explored in prior works. Such as PolyGen, Nexus.

Questions for authors
  1. Can the authors provide a more controlled comparison between joint and factorized flow models to demonstrate that the factorization itself simplifies learning and improves mesh quality?

  2. How does proposed method handle empty/solid bottles? this case study would help calrify when factorization is beneficial and where it may struggule.

  3. Could the authors clarify the novelty of the proposed factorization compared with prior works such as PolyGen and Nexus? In particular, what are the key methodological advances beyond applying flowing matching to existing two-stage framework?

  4. In qualitative results of Figure 5, if we repair broken surface regions from baseline methods via boundary triangulation or surface interpolation, how is the results compared with LATO.2?

2 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/2]

We sincerely thank the reviewer for the insightful feedback and constructive suggestions. We respectfuly explain that our core contribution lies in factorizing and coupling dedicated vertex and topology latent spaces for native mesh flow matching. This generative formulation directly enables scalable mesh synthesis, part-wise refinement, and topology-adaptive editing. Prior decoupled methods rely on either slow autoregressive decoding or deterministic prediction prone to topological redundancy, whereas our probabilistic latent flow design achieves both fast inference and clean connectivity. Thus, high-level conceptual precedents do not diminish the novelty or empirical significance of our framework.

Below, we isolate the architectural gains of our factorized latent framework, clarify how coarse voxel conditioning resolves topological ambiguity for identical vertex sets, distinguish LATO.2 from prior vertex-first approaches, and evaluate post-hoc boundary-based hole filling on the generated meshes.

W1 & Q1: Isolating the Contribution of Factorization

We agree that isolating factorization from resolution, offset prediction, and training data is critical.

  • Representational Efficiency & Parameter Budget: Because LATO entangles continuous vertex coordinates and discrete connectivity in a single joint latent space, its VAE is constrained to 1283128^3 encoding and 5123512^3 decoding. Decoupling these modalities allows LATO.2 to compress geometry into a much more compact 64364^3 latent while decoding to 102431024^3—using a smaller flow model (160M160\text{M} vs. 190M190\text{M} parameters in LATO).
  • Training Data Clarification: The flow models in LATO and LATO.2 are trained on the exact same dataset. Additional synthetic shapes (simple geometric primitives such as spheres and boxes with augmentation) were used exclusively as geometric augmentation when training the Vertex VAE.
  • Controlled Comparison (Table R1): Due to the limited rebuttal window, we could not retrain LATO.2 from scratch at 1283→5123128^3 \to 512^3 (which we will include in the final revision). To isolate factorization in the meantime, we evaluated an existing 643→512364^3 \to 512^3 LATO.2 checkpoint with the offset head disabled (using voxel centers directly), removing both the 102431024^3 resolution advantage and sub-voxel offset refinement. Even under a tighter 64364^3 latent bottleneck (vs. LATO's 1283128^3) and a smaller flow backbone (160M160\text{M} vs. 190M190\text{M}), LATO.2 @64-512 (no offset) still outperforms LATO @128-512 across every metric—most notably reducing Hausdorff Distance from 0.07380.0738 to 0.06790.0679. This confirms that factorization itself reduces topological holes and improves structural fidelity.

Table R1: Controlled ablation isolating latent factorization from decoding resolution, offset prediction, and model size.

MethodFlow Model SizeCD(L2)↓\text{CD}(L_2)\downarrowCD(L1)↓\text{CD}(L_1)\downarrowHD↓\text{HD}\downarrow∣NC∣↑\lvert\text{NC}\rvert\uparrow
LATO @128-512190M0.04210.06170.07380.8262
LATO.2 @64-1024160M0.04070.05960.06570.8341
LATO.2 @64-512 (no offset)160M0.04100.06000.06790.8311

W2 & Q2: Effectiveness of the vertex-conditioned topology generation

To clarify both the empirical efficacy of T-Flow and how it disambiguates identical vertex sets (such as empty versus solid bottles), we highlight two key aspects of our design:

  • Consistency & Robustness: Table 4 of the original paper evaluates T-Flow conditioned on both ground-truth vertices and V-Flow generated vertices, demonstrating strong robustness to vertex perturbations. Furthermore, when compared directly against a deterministic connectivity predictor (FastMesh) on identical ground-truth vertices (Table R2 and Figure 6 in the revision pdf), T-Flow yields substantially cleaner and more consistent topology.

  • Resolving Identical Vertices with Different Topologies (Empty vs. Solid Bottle): Crucially, T-Flow is not conditioned solely on raw vertex coordinates. As illustrated in Figure 2 of the main paper, it is also conditioned on the coarse 64364^3 voxel occupancy grid, the same initial scaffold provided to V-Flow, which supplies volumetric cues about interior partitions and surface boundaries.

  • Case Study (Figure 14, Appendix in the revision pdf, Page 21): We provide the exact bottle test case requested. Given the same set of bottle vertices, one 64364^3 voxel condition specifies an internal horizontal partition while the other does not; T-Flow accurately reconstructs the appropriate distinct topology for each condition.

  • Sub-Voxel Ambiguity: When two valid topologies differ only at a fine sub-voxel scale where the 64364^3 conditioning is identical, T-Flow draws a coherent sample from its learned conditional topology distribution.

AuthorsOpenReview ↗

Rebuttal by Authors [2/2]

W3 & Q3: Methodological Novelty Relative to PolyGen, FastMesh, and Nexus

While predicting connectivity after vertices has precedents in PolyGen and FastMesh (which we have credited explicitly in Line 257-263 in the revision pdf), LATO.2 differs fundamentally in its latent representation and continuous generative formulation:

  • Comparison with PolyGen and FastMesh: PolyGen relies on autoregressive decoding for both vertices and faces, incurring prohibitively slow inference. FastMesh pairs autoregressive vertices with direct feed-forward adjacency prediction; while fast, this deterministic predictor produces severe face redundancy (evidenced by its 5.625.62 face-to-vertex ratio and 65.00% non-manifold edges in Table R2).
  • LATO.2 Formulation: Rather than using autoregressive or deterministic feed-forward predictors, LATO.2 learns separate continuous latent spaces for vertices and topology. Connectivity is encoded via adjacency-masked attention and sampled through conditional flow matching over realized vertices and 64364^3 voxels—providing a learned, probabilistic topology distribution suited for generative sampling.
  • Comparison with Nexus: As noted around Line 105 of the main paper, the closest concurrent work, Nexus, uses a coarse-to-fine octree diffusion model that generates vertices level by level, making it considerably slower than LATO.2's single-pass structured latent decoding. For topology, Nexus uses graph autoencoding supervised by a SpaceTime loss, whereas LATO.2 trains its topology latent with a contrastive objective and extracts faces via loop detection. Moreover, Nexus has not publicly released its code; our open-source release will provide a reproducible benchmark for factorized mesh generation.

As shown in Table R2 (evaluated on identical input vertices), T-Flow outperforms both PolyGen and FastMesh in geometric accuracy (CD\text{CD} and HD\text{HD}), normal consistency (∣NC∣\lvert\text{NC}\rvert), Topology Score (TS\text{TS}), and manifoldness (NM%\text{NM}\%), while running over 100×100\times faster than PolyGen.

Table R2: Comparison with factorized vertex-and-topology generation methods on identical input vertices.

MethodCD(L1)×103↓\text{CD}(L_1)\times 10^3\downarrowCD(L2)×103↓\text{CD}(L_2)\times 10^3\downarrowHD×103↓\text{HD}\times 10^3\downarrow∣NC∣↑\lvert\text{NC}\rvert\uparrowTS↑\text{TS}\uparrowTime (s)↓\downarrowFace CountF/VNM (%)↓\downarrow
GT----0.723-25381.921.07
PolyGen16.682.7251.8480.7990.612174.523711.7914.32
FastMesh1.2740.0780.3730.9230.6190.24574355.6265.00
T-Flow (Ours)1.2470.0510.3680.9640.6341.33826532.016.68

Q4: Post-Hoc Boundary Triangulation / Hole Filling on Figure 5

Following your suggestion, we applied post-hoc boundary triangulation to the generated meshes of every method. Specifically, we detected closed loops formed by open boundary edges and triangulated each loop containing fewer than 3030 edges.

Table R3 reports the resulting metrics and deltas after hole filling. Overall, post-hoc triangulation yields negligible or slightly negative changes across baselines (particularly in ∣NC∣\lvert\text{NC}\rvert and TS\text{TS}) and fails to narrow the performance gap with LATO.2. This occurs for two reasons:

  1. Most broken regions observed in Figure 5 are not bounded by clean, closed loops of open boundary edges (often involving non-manifold or disconnected adjacent edges), so boundary triangulation repairs only a small fraction of defects.
  2. Blindly triangulating boundary loops frequently seals legitimate openings in valid structures, degrading normal consistency and topological regularity.

These results confirm that LATO.2’s advantages arise from intrinsically cleaner connectivity generation rather than surface artifacts that could be trivially patched in post-processing.

Table R3: Performance and metric deltas (Δ\Delta relative to raw outputs) after boundary-loop hole filling on generated meshes.

MethodTypeCD(L2)↓\text{CD}(L_2)\downarrowHD↓\text{HD}\downarrow∣NC∣↑\lvert\text{NC}\rvert\uparrowTS↑\text{TS}\uparrow
BPTAR0.0604 (+0.0001)0.1075 (+0.0008)0.8095 (-0.0016)0.7764 (-0.0003)
DeepMeshAR0.0530 (+0.0001)0.0964 (+0.0018)0.8192 (-0.0026)0.8036 (-0.0017)
MeshRippleAR0.0458 (-)0.0951 (+0.0010)0.8173 (-0.0001)0.7794 (-0.0002)
MeshFlowFlow0.0459 (+0.0004)0.0780 (+0.0008)0.8225 (-0.0002)0.6881 (+0.0001)
LATOFlow0.0424 (+0.0003)0.0726 (-0.0012)0.8220 (-0.0042)0.8050 (-0.0007)
LATO.2Flow0.0407 (-)0.0660 (+0.0003)0.8340 (-0.0001)0.8095 (-)
Reviewer_vhLwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

LATO.2 decomposes explicit mesh generation into two stages:

  1. V-Flow generates vertices by decoding a V-VAE to obtain coordinates with sub-voxel offsets.
  2. T-Flow generates topological latent variables based on the generated vertices and coarse geometric structures, then decodes edge probabilities and reconstructs faces. This work further leverage this decomposition to support vertex budget control, part-based generation, and regeneration of connectivity. Experiments across multiple settings demonstrate the effectiveness of the proposed method.
Critical strengths
  1. The decomposition of vertices and edges introduces an explicitly controllable intermediate representation of vertices, decoupling geometric editing from subsequent connectivity generation.
  2. The capabilities for editing and part-based generation are practically appealing.
  3. The method outperforms the listed baselines on multiple reported metrics.
Critical weaknesses
  1. The evaluation primarily reports Chamfer Distance (CD), Hausdorff Distance (HD), and normal consistency. While these metrics help assess the quality of the final surfaces, they do not independently establish improvements in topology.
  2. The edge-to-face decoding process is insufficiently explained. For example, it is unclear whether manifold constraints are enforced and how overlapping faces are handled.
  3. The results presented so far are relatively simple, and the paper does not sufficiently demonstrate the overall generation quality.
Questions for authors
  1. How does the method ensure that valid meshes can be reconstructed from predicted edges? How does it handle non-manifold structures, open boundaries, and degenerate faces?
  2. How do the authors demonstrate that the proposed method genuinely improves topology quality? Why do the main results focus primarily on CD, HD, and NC, rather than directly reporting topology-validity metrics, such as the proportion of quadrilateral faces and the number of faces?
  3. Does high-resolution generation offer genuine computational advantages? How does the method perform in terms of both computational efficiency and generation quality?
2 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/2]

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of our explicit intermediate representation, practical editing capabilities, and strong quantitative performance. We agree that evaluating topological quality directly and clarifying our decoding mechanics are essential. Below, we address your concerns regarding topological evaluation, edge-to-face decoding, visual complexity, and computational efficiency point by point.

W1 & Q2: Direct Evaluation of Topological Quality

Because no standardized topology metric has yet been universally adopted in the community, recent baselines such as Nexus (SIGGRAPH 2026) and MeshFlow (CVPR 2026) primarily report geometric metrics. To rigorously verify LATO.2’s topological improvements, including wireframe regularity and quad-mergeable structure, we evaluate two additional topology-focused measures across all baselines:

  1. Topology Score (TS): Following MeshRFT, we compute TS, which evaluates meshes after triangle-to-quad merging on angle quality, aspect ratio, and consistency between adjacent quadrilaterals.
  2. Human User Study: We conducted a preference study specifically assessing wireframe and topological quality alongside geometry across 32 evaluation cases.

As shown in Tables R1 and R2, LATO.2 achieves the highest TS (0.80950.8095) and obtains a 48.23% user preference ratio, more than three times that of the next best method. These results confirm that our factorized formulation genuinely improves structural topology in addition to geometric accuracy.

Table R1: Comparison of Topology Scores (TS) and geometric metrics

MethodTypeCD(L2)↓\text{CD}(L_2)\downarrowHD↓\text{HD}\downarrowTS↑\text{TS}\uparrow
BPTAR0.06030.10670.7767
DeepMeshAR0.05290.09460.8053
MeshRippleAR0.04580.09410.7796
MeshFlowFlow0.04550.07720.6880
LATOFlow0.04210.07380.8057
LATO.2Flow0.04070.06570.8095

Table R2: User study preference ratio (wireframe topology and geometric balance).

MethodTypeUser Preference Ratio$\uparrow$
BPTAR0.1285
DeepMeshAR0.0521
MeshRippleAR0.1467
MeshFlowFlow0.0563
LATOFlow0.1341
LATO.2Flow0.4823

W2 & Q1: Edge-to-Face Decoding, Manifoldness, and Boundary Handling

Thanks for pointing out this lack of clarity. We have expanded Section 3 in the revision pdf (Line 241-243) to detail our edge-to-face reconstruction process and its topological properties:

  • 3-Loop Extraction & Degenerate Faces: We recover faces from the predicted adjacency graph via 3-loop detection: whenever three predicted edges form a v1v_1–v2v_2–v3v_3–v1v_1 cycle, they define a triangular face. Loop detection is a reliable and widely used heuristic in generative mesh pipelines (e.g., MeshFlow [CVPR 2026] uses 4-loops for quad extraction). Because any valid 3-loop requires three distinct vertices (v1≠v2≠v3v_1 \neq v_2 \neq v_3), zero-area degenerate faces with repeated vertex indices are structurally prevented.
  • Non-Manifold Structures, Open Boundaries, and Overlapping Faces: Unlike implicit field-based methods (e.g., SDFs) that force outputs to be watertight, our graph formulation does not impose hard post-hoc manifold constraints. Instead, it natively represents open boundaries (edges shared by a single 3-loop) and valid non-manifold structures (edges shared by ≥3\ge 3 loops) directly from the artist-mesh training distribution. Because T-Flow learns clean local adjacency through a contrastively regularized latent space, spurious overlapping faces and extraneous edges are rare in practice, as reflected by our leading TS and human preference scores (Tables R1 and R2).
  • Robustness to Noisy Vertices: As demonstrated in Table 4 of the main paper, T-Flow remains robust under vertex-stage perturbations, avoiding catastrophic topological failure even when conditioned on noisy vertex inputs.
AuthorsOpenReview ↗

Rebuttal by Authors [2/2]

W3: Visual Complexity and Overall Generation Quality

We appreciate the suggestion to showcase more complex assets and clearer visual evidence. Figure 13 in the Appendix in the original provides a broader gallery of intricate meshes and part-wise generation results. In addition, we have added multi-view visualizations in Figure 15 in the revision pdf (Page 22) so that fine geometric details, surface completeness, and wireframe connectivity can be inspected across viewpoints. Together, these visual results complement the quantitative benchmarks in Table 1 of the original paper and better demonstrate LATO.2’s overall generation fidelity.

Q3: Computational Efficiency and Quality of High-Resolution Generation

LATO.2 achieves genuine computational savings by decoupling the generative latent resolution from the geometric decoding resolution:

  1. Compact Latent Flow Matching (64364^3 vs. 1283128^3): In LATO.2, the V-VAE compresses vertex geometry into a compact 64364^3 volumetric latent space. Flow matching operates exclusively at this 64364^3 resolution, compared to 1283128^3 in the original LATO, yielding more than a 2×2\times speedup in 24-step sampling time and a 3×3\times–4×4\times reduction in token count (Table R3).
  2. Linear-Scaling Sparse Decoding to 102431024^3: High-resolution vertex coordinates (102431024^3) are recovered entirely during deterministic decoding via hierarchical sparse subdivision (643→1024364^3 \to 1024^3). By pruning empty voxels at each stage, only occupied surface voxels are retained. Consequently, token counts scale roughly linearly with the actual vertex count rather than cubically with grid volume. As shown in Table R4, each sparse decoder layer takes only 2828–33 ms33\text{ ms} on average, keeping total decoding overhead around ∼120\sim 120–200 ms200\text{ ms} per mesh.
  3. Genuine Quality Gains at 102431024^3: Finally, Table R5 compares full LATO.2 (102431024^3) against a quantized 5123512^3 baseline (LATO.2 q512), confirming consistent improvements across all geometric metrics—particularly in Hausdorff Distance and Normal Consistency. Thus, LATO.2 delivers 102431024^3 geometric precision at the generative cost of a 64364^3 latent flow.

Table R3: Flow-matching sampling cost on held-out artist meshes (64 meshes).

Latent ResolutionMean / Max Input TokensMean / Max Sampling Time @24 steps (s)
64364^3 (LATO.2)9,722 / 35,4593.0 / 11.0
1283128^3 (LATO)32,665 / 151,5726.3 / 25.1

Table R4: Per-layer sparse decoding cost on held-out artist meshes (64 meshes).

Decoder LayerMean / Max Input TokensMean / Max Time (ms)
Surface2Vertex9,722 / 35,4590.77 / 1.35
64→12864 \to 128959 / 2,56233.0 / 64.0
128→256128 \to 2561,183 / 2,95129.5 / 59.0
256→512256 \to 5121,320 / 3,15729.5 / 61.7
512→1024512 \to 10241,325 / 3,15728.3 / 58.3

Table R5: Geometric quality comparison at 5123512^3 vs. 102431024^3 decoding resolution.

MethodCD(L2)↓\text{CD}(L_2)\downarrowCD(L1)↓\text{CD}(L_1)\downarrowHD↓\text{HD}\downarrow∣NC∣↑\lvert\text{NC}\rvert\uparrow
LATO.2 q5120.04100.06000.06790.8311
LATO.2 (102431024^3)0.04070.05960.06570.8341
Reviewer_uL6AOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LATO.2, which splits explicit mesh generation into two flow-matching stages: a vertex flow that generates vertex positions and a topology flow that predicts connectivity from those vertices. The split enables part-wise generation at higher resolution and re-connection after vertex edits, and the paper reports lower CD/HD and higher normal consistency than AR and flow-based baselines.

Critical strengths
  • The factorized design makes the generated connectivity editable: after vertices are moved, added or merged, re-running T-Flow gives consistent faces. Fig. 4 and Fig. 9 show convincing repairs of tearing and seams. This is a practical capability for downstream asset editing. (However, predicting connectivity from given vertices is not new. A similar idea appears in PolyGen (ICML 2020), SpaceMesh (SIGGRAPH Asia 2024) and FastMesh (3DV 2026). The reported results are better, but the gain may come from other factors, such as higher vertex resolution and a larger training set. A fair ablation is needed to support this contribution; see W2.)
Critical weaknesses
  1. Novelty claims. (a) The factorized framework. Splitting vertices from connectivity is the key change over LATO, but the idea itself is not new. PolyGen (ICML 2020) and FastMesh (3DV 2026) already predict faces from generated vertices. The contribution is therefore bringing this known split into a latent flow framework, and the benefit of doing so is not yet shown (see W2).

(b) The two modules. Both are small extensions of existing techniques. Regressing a continuous vertex position inside each voxel cell is standard (Neural Dual Contouring (SIGGRAPH’22)). Encoding connectivity as per-vertex embeddings, predicted from given vertex positions, is the core of SpaceMesh (SIGGRAPH Asia 2024). What remains new is sampling these per-vertex features with a flow. However, no fair comparison is given to show this sampling actually improves connectivity over deterministic prediction (see W2).

(c) Better results and new applications. The gain over LATO is small (~3%) and it may come from higher resolution and more training data. The abstract also calls part-wise generation “unique to this factorization”, which is overstated. Part-wise generation does not depend on the factorization: FullPart (ICLR 2026), MeshMosaic (CVPR 2026) and PartDiffuser (CVPR 2026) already generate shapes part by part. None of these are discussed. The only factorization-specific step is stitching parts by re-running T-Flow, and its seam quality is not evaluated.

  1. The benefit of factorization is not isolated. The paper’s central claim is that a joint latent causes drifting vertices and broken surfaces, and that factorization fixes this. This claim is never tested directly. Compared with LATO, the method changes several things at once: 1024³ instead of 512³ resolution, a new offset head, and 100K extra synthetic training meshes. Any of these could explain the gain. Two controlled ablations are needed: (i) a joint latent vs. a factorized latent with the same backbone, resolution and data; (ii) T-Flow vs. a deterministic vertex-to-face predictor on the same vertices. Sec. 4.3 states that (ii) was done, but no numbers are reported.

  2. Topology quality is not verified. Cleaner connectivity is the paper’s main selling point, yet it reports only CD, HD and NC, which barely reflect topology. Faces are recovered by loop detection, which does not guarantee a manifold mesh. Errors in the vertex stage also cannot be corrected later and may lead to spurious faces or wrong connections between nearby surfaces. Neither risk is examined.

Questions for authors

Please address W1–W3, especially the controlled ablations in W2 and the topology evaluation in W3.

4 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/4]

We sincerely thank the reviewer for the thoughtful assessment, constructive feedback, and recognition of our framework’s practical editing and repair capabilities. While the concept of vertex and connectivity decoupling, and several individual components have precedents in prior work, our core contribution lies in factorizing and coupling dedicated vertex and topology latent spaces for native mesh flow matching. This generative formulation directly enables scalable mesh synthesis, part-wise refinement, and topology-adaptive editing. Prior decoupled methods rely on either slow autoregressive decoding or deterministic prediction prone to topological redundancy, whereas our probabilistic latent flow design achieves both fast inference and clean connectivity. Thus, high-level conceptual precedents do not diminish the novelty or empirical significance of our framework.

Below, we address each concerns in detail and provide new controlled ablations and topology evaluations.

W1: (a) Novelty of the Factorized Framework

We agree that generating vertices prior to connectivity has important precedents in PolyGen and FastMesh, and we have discussed both works explicitly in Section 3.2 (Line 257-263) in the revised pdf. However, LATO.2 differs substantially in both its underlying representation and its generative formulation:

  • Comparison with PolyGen and FastMesh: PolyGen relies on an autoregressive formulation for both vertex coordinates and face indices, resulting in prohibitively slow sequential inference. FastMesh pairs autoregressive vertex generation with direct feed-forward adjacency prediction; although much faster, deterministic feed-forward prediction lacks distributional modeling and generates severe face redundancy (evidenced by its inflated face-to-vertex ratio and high non-manifold edge rate in Table R1).

  • LATO.2 Formulation: Rather than operating autoregressively or deterministically in explicit discrete space, LATO.2 learns separate continuous latent spaces for vertex placement and mesh connectivity. Connectivity is encoded via adjacency-masked attention and sampled through a conditional flow model conditioned on realized vertices, thereby modeling a factorized conditional topology distribution amenable to continuous generative sampling.

  • Comparison with Concurrent Work (Nexus): As discussed around Line 105 of the original paper, the closest concurrent work, Nexus, employs a coarse-to-fine octree diffusion model that generates vertices level by level, whereas LATO.2 compresses vertices into a structured latent decoded in a single pass. For connectivity, Nexus formulates topology learning as graph autoencoding supervised by a SpaceTime loss, whereas LATO.2 trains a topology latent via a contrastive objective and reconstructs faces through loop detection.

To demonstrate the concrete benefits of bringing this factorization into a latent flow framework, Table R1 compares PolyGen, FastMesh, and our T-Flow under the exact same input vertex set. T-Flow achieves lower Chamfer and Hausdorff distances, higher normal consistency (|NC|) and Topology Score (TS), a face-to-vertex ratio (F/V\text{F/V}) closely aligned with Ground Truth (2.012.01 vs. 1.921.92), and dramatically cleaner manifoldness (6.68% non-manifold edges vs. 65.00% for FastMesh and 14.32% for PolyGen), while remaining nearly two orders of magnitude faster than PolyGen. These results have been incorporated into Section 4.1(Line 456-463; Table 2) of the revised paper.

Table R1: Quantitative comparison of connectivity generation methods given the same input vertex set.

MethodCD(L1)x10^3↓CD(L2)x10^3↓HDx10^3↓|NC|↑TS↑Time(s)↓Face CountF/VNM(%)↓
Gt----0.723-25381.921.07
PolyGen16.682.7251.8480.7990.612174.523711.7914.32
FastMesh1.2740.0780.3730.9230.6190.24574355.6265.00
T-Flow(Ours)1.2470.0510.3680.9640.6341.33826532.016.68
AuthorsOpenReview ↗

Rebuttal by Authors [2/4]

W1: (b) Distinctions of the Vertex and Topology Modules

We thank the reviewer for highlighting the connections to Neural Dual Contouring (NDC) and SpaceMesh. While we draw inspiration from these representations, both modules in LATO.2 differ fundamentally in formulation and generative role:

  • Vertex Module vs. NDC: NDC predicts within-cell vertex locations and extracts surface connectivity via deterministic dual-contouring rules tied directly to the grid. In contrast, our voxel scaffold serves as an anchor for sparse latent vertex generation, completely decoupled from connectivity rules. The offset head acts solely as a sub-voxel precision refinement component rather than a standalone surface reconstruction primitive.

  • Topology Module vs. SpaceMesh: SpaceMesh deterministically regresses adjacency and edge-ordering embeddings directly from vertex coordinates and local geometric context. Conversely, LATO.2 learns an adjacency conditioned topology autoencoder paired with a vertex-conditioned flow over its regularized latent space, explicitly separating connectivity encoding, conditional latent sampling, and decoding. Our contribution here lies in this continuous conditional generative formulation rather than the general use of per-vertex features. We have included these attributions and distinctions explicit in the revised pdf in Section 3.1(Line 221-226).

  • Value of Flow-Based Sampling over Deterministic Prediction: As demonstrated under identical input vertices in Table R1, probabilistic sampling via T-Flow avoids the catastrophic over-prediction and topological artifacts produced by deterministic feed-forward predictors (reducing non-manifold edges from FastMesh’s 65.00% down to 6.68%, and F/V\text{F/V} from 5.625.62 to 2.012.01).

W1(c): Scope of Part-Wise Generation and Evaluation of Seam Quality

We appreciate the reviewer’s point regarding prior part-based generation methods (FullPart, MeshMosaic, and PartDiffuser) and acknowledge that our phrasing in the Abstract was overly broad. Our intended statement is the part-wise vertex generation and global connectivity stitching is unique to this factorization. While those prior works operate within field-based or autoregressive mesh generation frameworks, to the best of our knowledge, LATO.2 is the first to enable part-wise synthesis within a flow-based native mesh generation paradigm. Specifically, what our factorization uniquely enables is a part-wise vertex generation, assembly, and global connectivity stitching mechanism: each part is normalized and decoded at 102431024^3 resolution to utilize the model's full spatial capacity, and the assembled parts are then seamlessly connected by T-Flow. We have narrowed and clarified this claim in the Abstract and Introduction, and added a discussion of FullPart, MeshMosaic, and PartDiffuser in Section 3.3 ("Part-wise Generation and Refinement", Line 344-348).

Seam QualityTo quantitatively evaluate seam quality, we remove one or more rings of faces along the junctions between mesh parts on held-out artist-created meshes and compare MeshMosaic against T-Flow in reconstructing the missing seam faces. As reported in Table R2 in below (with qualitative close-ups in Figure 9 of the Appendix in the revised pdf), T-Flow outperforms MeshMosaic across every metric—halving CD(L2)\text{CD}(L_2) and non-manifold edge rate while matching the ground-truth face count much more closely.

Table R2: Comparison of seam quality on held-out artist meshes.

MethodCD(L2)x10^3↓HDx10^3↓|NC|↑Face CountNM(%)↓
Gt---55.42.54
MeshMosaic1.08063.070.96861.46.86
Ours0.514947.400.97355.23.07
AuthorsOpenReview ↗

Rebuttal by Authors [3/4]

W2 (and W1c): Isolating the Architectural Gains of Factorization

We fully agree on the importance of isolating the effect of factorization from confounding factors such as resolution, the offset head, and training data.

(i) Joint Latent vs. Factorized Latent (Controlled Comparison)

  • Inherent Representational Capacity: Because LATO entangles continuous vertex positions and discrete connectivity in a single joint latent space, its VAE is bottlenecked at encoding 1283128^3 voxels and decoding 5123512^3. By decoupling these two heterogeneous modalities, LATO.2 achieves higher spatial compression (64364^3 encoded voxels) while supporting 102431024^3 decoding. This structural efficiency contributes directly to the cleaner edge layouts and reduced hole artifacts shown in Figure 5 of the main paper.
  • Training Data Clarification: We respectfully clarify that the flow models in LATO and LATO.2 are trained on the exact same mesh dataset. The additional synthetic shapes (simple geometric primitives such as spheres and boxes with augmentations) were used solely as geometric augmentation for training the Vertex VAE.
  • Controlled Ablation (Table R3): Due to the limited rebuttal window, we could not retrain LATO.2 from scratch at the 1283→5123128^3 \to 512^3 configuration (we commit to adding that exact matched run in future revision). In the interim, we evaluated an existing LATO.2 checkpoint trained at 643→512364^3 \to 512^3 and disabled the offset head (using voxel centers directly) to eliminate both the 102431024^3 resolution advantage and the offset refinement. Even under a strictly tighter latent bottleneck (64364^3 vs. LATO's 1283128^3) and without offsets, LATO.2 @64-512 (no offset) consistently outperforms LATO @128-512 across all metrics, most notably in Hausdorff Distance (0.06790.0679 vs. 0.07380.0738), demonstrating that the factorized formulation itself mitigates drifting vertices and topological holes.

Table R3: Controlled ablation isolating the effect of latent factorization from decoding resolution and the offset head.

MethodCD(L2)↓\text{CD}(L_2)\downarrowCD(L1)↓\text{CD}(L_1)\downarrowHD↓\text{HD}\downarrow∣NC∣↑\lvert\text{NC}\rvert\uparrow
LATO @128-5120.04210.06170.07380.8262
LATO.2 @64-10240.04070.05960.06570.8341
LATO.2 @64-512 (no offset)0.04100.06000.06790.8311

(ii) T-Flow vs. Deterministic Vertex-to-Face Prediction

We apologize for the omission in Section 4.3. While Table 4 of the main paper ablates T-Flow’s robustness across different input sources and configurations, the direct quantitative comparison against a deterministic vertex-to-face predictor (FastMesh) on identical vertices is now explicitly provided in Table R1 above. As shown, T-Flow improves CD(L2)\text{CD}(L_2) from 0.0780.078 to 0.0510.051, |NC| from 0.9230.923 to 0.9640.964, and reduces non-manifold edges from 65.00% to 6.68%. We will ensure these numbers are reported and cross-referenced clearly in Sections 4.1 and 4.3 of the revised paper.

AuthorsOpenReview ↗

Rebuttal by Authors [4/4]

W3: Topology Validation and Loop Detection

We agree that geometric metrics (CD\text{CD}, HD\text{HD}, NC\text{NC}) alone do not adequately capture connectivity quality. Because there is not yet a single universally adopted metric for mesh topology in the community, prior works such as Nexus and MeshFlow typically omit topology-specific evaluations. To rigorously validate LATO.2's topological quality, we provide three complementary evaluations:

  1. Non-Manifoldness & Face Density (Tables R1 & R2): As shown in Tables R1 and R2, our method achieves substantially lower non-manifold edge percentages (NM%\text{NM}\%) and more faithful face counts (F/V\text{F/V}) than competing connectivity predictors.

  2. Topology Score (Table R4): Following MeshRFT, we evaluate the Topology Score (TS) across all baselines on the Mixed test set. As shown in Table R4, LATO.2 achieves the highest TS (0.80950.8095) alongside the lowest CD(L2)\text{CD}(L_2) and HD\text{HD}.

  3. Human User Study (Table R5): We conducted a preference study across 3232 test cases, asking participants to select the generated mesh offering the best joint balance of geometry and topology. LATO.2 was preferred in 48.23% of cases—more than three times as often as the strongest baseline.

Table R4: Comparison of generation quality on the Mixed test set.

MethodTypeCD(L2)↓HD↓TS↑
BPTAR0.06030.10670.7767
DeepMeshAR0.05290.09460.8053
MeshRippleAR0.04580.09410.7796
MeshFlowFlow0.04550.07720.6880
LATOFlow0.04210.07380.8057
LATO.2Flow0.04070.06570.8095

Table R5: Human preference ratio across 32 evaluation cases (best balance of topology and geometry).

MethodTypeUserStudy↑
BPTAR0.1285
DeepMeshAR0.0521
MeshRippleAR0.1467
MeshFlowFlow0.0563
LATOFlow0.1341
LATO.2Flow0.4823

On Loop Detection and Sensitivity to Vertex-Stage Errors: Loop detection is a well-established and effective mechanism for face recovery in recent generative pipelines (for example, MeshFlow [CVPR 2026] similarly utilizes 4-loops for face reconstruction). Although loop detection does not enforce strict manifoldness by construction, our contrastively trained topology latent yields clean edge graphs in which spurious connections and non-manifold artifacts are rare in practice (6.68% in Table R1 and 3.07% in Table R2). Furthermore, as shown in Table 4 of the main paper, T-Flow generalizes reliably across vertices generated from diverse sources, demonstrating strong tolerance to minor vertex-stage perturbations. We have added an explicit discussion of manifoldness, vertex-error sensitivity in the limitations section (Line 1058-1063) the revised pdf.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Authors’ Summary of Revisions and Responses to Reviewer Concerns

We thank all reviewers for their constructive feedback. Below, we summarize the manuscript revisions grouped by theme. Line numbers refer to the revised PDF; W and Q denote the corresponding reviewer's weaknesses and questions. The revised passages are highlighted in blue.

1. Clarification of Claims and Related Part-Wise Work

  • Section 1, Introduction (lines 111–113) — uL6A, W1(c): Narrowed the uniqueness claim to the mechanism of part-wise vertex generation, vertex assembly, and subsequent connectivity generation with T-Flow, while retaining topology-adaptive editing.

  • Section 3.3, Part-wise Generation and Refinement paragraph (lines 344–349) — uL6A, W1(c): Added a discussion of FullPart, MeshMosaic, and PartDiffuser, distinguishing their part-generation mechanisms from our part-wise vertex generation, assembly, and T-Flow connectivity generation.

2. Methodological Distinctions from SpaceMesh, PolyGen, and FastMesh

  • Section 3.1, Vertex and Topology VAE (lines 221–226) — uL6A, W1(b): Added a comparison with SpaceMesh, clarifying that our contribution is learning and sampling a conditional topology latent distribution through a topology VAE and vertex-conditioned flow, rather than using per-vertex embeddings alone.

  • Section 3.2, Factorized Mesh Generation (lines 257–263) — uL6A, W1(a); 4hXt, W3/Q3: Credited PolyGen and FastMesh as precedents for generating vertices before connectivity, and clarified the distinction of our separate continuous latent spaces and conditional flow formulation.

3. Topological Evaluation and Controlled Connectivity Comparisons

  • Section 4.1, Geometry Conditioned Mesh Generation paragraph (lines 422–425; Table 3) — uL6A, W3; vhLw, W1/Q2: Added a user study over 32 cases evaluating the joint balance of topology and geometry, with LATO.2 receiving a 48.23% preference ratio.

  • Section 4.1, Topology Generation paragraph (lines 456–463; Table 2) — uL6A, W2(ii)/W3; vhLw, W1/Q2; 4hXt, W2/Q2: Added comparisons with PolyGen and FastMesh on identical input vertices, reporting geometric fidelity, runtime, face counts, face-to-vertex ratios, and non-manifold edge percentages.

  • Section 4.2, Topology Generation paragraph (lines 473–476; Figure 6) — uL6A, W2(ii); vhLw, W3; 4hXt, W2/Q2: Added qualitative comparisons with PolyGen and FastMesh to illustrate differences in local geometric detail and surface connectivity.

4. Part-Wise Seam Evaluation, Conditioning Case Study, and Multi-View Visualizations

  • Appendix A.4, Seam Quality in Part-wise Generation (lines 909–932; Table 8 and Figure 9) — uL6A, W1(c): Added quantitative and qualitative comparisons with MeshMosaic on reconstructing missing seam faces after removing rings of faces at part junctions.

  • Appendix A.6, Topology conditioning paragraph (lines 1024–1025 and 1047–1051; Figure 14) — 4hXt, W2/Q2: Added a bottle case study using identical vertices and different coarse voxel conditions, showing connectivity generation with and without an internal horizontal surface.

  • Appendix A.6, Multi-view visualizations paragraph (lines 1053–1055; Figure 15) — vhLw, W3: Added multiple views of complex generated meshes to support closer inspection of geometric details and local connectivity.

5. Face Reconstruction, Manifoldness, and Vertex-Error Sensitivity

  • Section 3.1, Vertex and Topology VAE (lines 241–243) — uL6A, W3; vhLw, W2/Q1: Specified the face-recovery rule: three distinct vertices form a triangular face when all three connecting edges are present in the predicted adjacency graph (3-loop detection).

  • Appendix A.7, Discussion, Limitations paragraph (lines 1058–1063) — uL6A, W3; vhLw, W2/Q1: Explicitly discussed the absence of a manifoldness guarantee and the sensitivity of connectivity generation to vertex errors or missing vertices, including incorrect connections between nearby surfaces.

Area_Chair_baQzOpenReview ↗

Looking for reviewers' reply!!!

Dear Reviewers,

The authors have posted their rebuttal. Please read it and reply in the discussion, covering:

Whether each of your questions and concerns has been addressed. Whether your rating and recommendation change, and why. Any remaining concern that must be settled before the meta-review. Best.

Dear Reviewer,

Thank you again for your time and feedback. We would like to follow up on our response to your review. We would appreciate your feedback on whether our response addresses your concerns and whether there are any remaining points we could clarify. We would be happy to provide further information to support the discussion.

Best regards, The Authors

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

Related papers

Discussion (0)

Sign in to comment.