acceptodds
Under review as a conference paper at ICLR 2027

MeshOctave: Vertex Split-and-Rewire Cascades for Native Mesh Generation

12 views

Abstract

Generating compact, artist-style meshes with explicit topology typically relies on autoregressive models which incur prohibitive sequential per-token costs, or continuous flow models that depend on heuristic connectivity decoders. Next-scale generation paradigms offer a compelling alternative by enabling parallel intra-scale token prediction and coarse-to-fine refinement from global structure to local topology; yet, existing methods derive hierarchical scales via progressive mesh simplification and invert them sequentially. This eliminates intra-scale parallelism and scales generation steps linearly with face count. In this paper, we propose **MeshOctave**, which instead defines scale through dyadic spatial grid resolutions, framing coarsening as a deterministic collapse that merges vertices sharing a voxel cell and inherits connectivity. Its inverse operation, **split-and-rewire**, determines which octant sub-vertices are instantiated for each coarse face and resolves local connectivity using discrete structural tokens. These per-face operations require no serialization, each scale transition is modeled as an unordered set that adds one bit of coordinate precision, naturally supporting dynamic-length meshes and adaptive resolution refinement. We construct a scale-conditioned masked-uniform discrete diffusion model to learn split-and-rewire operation from resolution collapse hierarchies. MeshOctave outperforms strong baselines in geometric fidelity and topological validity by a non-trivial margin, while supporting adaptive resolution refinement and extending naturally to mesh subdivision tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3d generation, mask diffusion, artist mesh

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

This paper proposes MeshOctave, a generative model that predicts next-scale topology for native triangle meshes. MeshOctave defines scales by Octree grid resolution: halving the grid merges the vertices that share a voxel (“resolution collapse”), which gives a deterministic, globally parallel hierarchy. The corresponding inverse step, split-and-rewire, is encoded per coarse face as nine discrete tokens (child occupancy plus intra- and inter-parent adjacency) and learned with a scale-conditioned masked-uniform discrete diffusion model. The paper reports better geometric fidelity than autoregressive, coarse-to-fine, and flow-matching baselines. Furthermore, it demonstrates ~10× speedup over recent methods, and support for LOD generation and adaptive subdivision.

Critical strengths

1. Canonical, parallel hierarchy: The proposed resolution collapse is deterministic and fully parallel, and the number of generation steps grows with log(resolution) rather than raw face count. This seems to be a proper solution that fixes the sequential bottleneck of simplification-based hierarchies (VertexRegen, ARMesh).

2. Effective tokenization of topology: The proposed topology tokenization strategy, i.e., octant occupancy plus adjacency matrices, combine octree-style next-scale prediction with explicit connectivity. The design seems to be effective, where it needs no serialization and adapts the sequence length to shape complexity.

3. Strong visual results with ablations: The proposed MeshOctave consistently beats a broad set of recent baselines, and each design choice (discrete diffusion, uniform pathway, autoregressive connectivity head, 3D RoPE) is ablated.

Critical weaknesses

1. Inconsistent vertex/edge predictions may not be considered: While the tokens are per face, the vertices and edges are shared across faces, so the same occupancy and adjacency could be predicted several times. The paper does not provide details on how conflicting predictions are resolved, or how faces are formed from the decoded edges. Both are central to the claimed topological validity.

2. Inconsistent performance reports: The “Full” row in Table 3 (CD-L2 0.0430, HD 0.0739) differs from Table 2 (0.0392, 0.0645). With the numbers from Table 3, the proposed MeshOctave would fall below LATO.2 (0.0406). This reviewer thinks that this needs clarification.

3. Limited speed comparison: The speedup of the proposed method is measured only against two slow coarse-to-fine baselines, not against flow-matching or efficient AR methods, while the authors admit MeshOctave is slower than single-scale flow models. Furthermore, since the number of tokens could scale with face count, “independent of face count” claim might be overstated.

4. Under-specified evaluation: The test-set construction and train/test overlap are unclear from the paper. Crucially, only point-based geometric metrics are reported, with no topology or mesh-quality metrics (manifoldness, watertightness, degenerate faces).

Questions for authors
  1. (Relates to weakness 1) How are conflicting per-face predictions for shared vertices and edges merged, and how often do they disagree? What is the exact face-formation rule, and can it yield non-manifold faces?

  2. (Relates to weakness 2) Could the authors explain the number gap between the Table 2 and Table 3 “Full” results?

  3. (Relates to weakness 3) Could the authors report wall-clock time and memory against all possible baselines as a function of face count?

  4. (Relates to weakness 4) Please clarify the test set and report topology and mesh-quality metrics. Did authors use official checkpoints for the other baselines? How do the results change under matched data and compute?

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

We sincerely thank the reviewer for the constructive and detailed feedback. We appreciate your recognition of our deterministic hierarchy as a proper solution to the sequential bottleneck of simplification-based methods, as well as your positive remarks on our effective topology tokenization and strong visual results with ablations. Below, we address your concerns regarding conflicting predictions, performance discrepancies, runtime scaling, and evaluation details point by point.

W1 & Q1: Conflicting Predictions on Shared Elements and Face Construction

Thank you for raising this critical point. Because our tokens are defined per parent face, shared parent vertices and edges are indeed predicted multiple times across incident faces. However, within each scale, our mask pathway proposes structures, the uniform pathway revises earlier predictions, and confidence-based re-masking revisits uncertain tokens prior to advancing. Together, these mechanisms suppress cross-face disagreements substantially.

Empirically, across 200 validation samples, only 0.002%, 0.060%, and 0.001% of shared O\mathbf{O} vectors, S\mathbf{S} matrices, and C\mathbf{C} matrices exhibit conflicting predictions prior to post-processing (Table 1, measured at the complete vector/matrix level). After the final denoising step at each scale, any remaining conflicts are resolved via majority voting over complete vectors/matrices with deterministic tie-breaking; for inter-parent connectivity C\mathbf{C}, matrices are aligned to a canonical parent-edge orientation before voting and restored afterward.

Table 1. Cross-face inconsistency (%; ↓). O: occupancy; S/C: intra-/inter-parent connectivity.

MetricMasked DiffusionMask-Uniform
Occupancy (O)1.2210.002
Intra-parent connectivity (S)9.8190.060
Inter-parent connectivity (C)3.4510.001

Face construction and manifoldness. Each parent triangle has at most 24 candidate child vertices. After occupancy masking, we assemble the S\mathbf{S} and C\mathbf{C} blocks into a local undirected adjacency matrix A\mathbf{A}. For each triple of occupied vertices i<j<ki < j < k, we emit a triangle if and only if all three edges are present, that is, AijAjkAik=1A_{ij}A_{jk}A_{ik} = 1. Consequently, this face-formation rule can produce both open boundary edges and non-manifold edges (shared by more than two triangles). This is intentional for native mesh generation: our training corpus includes open surfaces and non-manifold assets, with no requirement that meshes be watertight or strictly manifold. We will make both the conflict-resolution procedure and the face-formation rule explicit in the revised manuscript.

W2 & Q2: Discrepancy Between Table 2 and Table 3 ("Full" Results)

Thank you for catching this discrepancy. The "Full" row in Table 2 of the paper reports our final full-scale model (624M parameters, trained for 56 GPU-days), which achieves 0.0392 CD and 0.0645 HD (surpassing LATO.2's 0.0406 and 0.0665). In contrast, to enable controlled and affordable comparisons across all ablation variants under a uniform compute budget, Table 3 uses a smaller "Full" backbone (276M parameters, trained for 24 GPU-days) shared across every row in the ablation study. Table 2 below details both configurations, and we will clearly distinguish them in the revision.

Table 2. Model configurations and training budgets for the "Full" rows in Paper Table 2 and Table 3. HG: one-sided hourglass depths; Blks: total Transformer blocks; Params: parameters (in millions).

Full modelWidthHGBlksHeadsFFNParamsGPUsDaysGPU-days
Paper Tab. 21024[2,4,6]241628166248756
Paper Tab. 3768[2,4,3]181220482768324
AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

W3 & Q3: Time and memory versus face count

Thank you for this helpful suggestion.

Scaling with output face count. In Revised Fig. 11, we extend our scaling evaluation to report both wall-clock inference time and peak allocated GPU memory as functions of output face count across all evaluated baselines. Table 3 in below summarizes these curves by output face-count interval, reporting the median runtime and peak memory within each bin. Thanks to intra-scale parallel decoding, MeshOctave exhibits substantially gentler runtime scaling than autoregressive and next-scale baselines such as BPT, Mesh-Silksong, DeepMesh, MeshRipple, ARMesh, and VertexRegen. Across approximately 0.5k–9k output faces, our runtime ranges from 30 to 105 s and peak allocated GPU memory ranges from 2.4 to 4.6 GiB. While latent-space flow methods (MeshFlow and LATO.2) achieve lower latency, MeshOctave attains better geometric fidelity (CD/HD) and substantially higher mesh quality across all MeshRFT metrics (Table 5), while directly modeling explicit vertex subdivision and connectivity to support flexible resolution control and mesh refinement.

Table 3. Time and peak memory versus output face count (summary of Fig. 11). Each cell reports the median time (s) / peak allocated GPU memory (GiB) for outputs in that face-count bin.

Method0–2k faces2–4k faces4–9.5k faces
MeshAnythingV252.1 / 3.81102.3 / 4.72—
BPT101.1 / 2.07180.0 / 2.58—
DeepMesh29.3 / 5.34167.3 / 5.73280.2 / 5.73
Mesh-Silksong43.9 / 1.68195.3 / 2.59306.7 / 3.23
MeshRipple107.9 / 4.62295.5 / 4.63729.0 / 4.65
MeshFlow4.4 / 6.575.6 / 6.57—
LATO.27.2 / 5.24—8.4 / 5.48
ARMESH193.0 / 3.05583.5 / 3.82—
VertexRegen372.3 / 4.24664.5 / 5.09—
Ours30.3 / 2.3945.7 / 3.1084.8 / 4.39

Revision of phrasing. We agree that describing inference cost as independent of face count is imprecise, as sequence length still grows with surface complexity at finer scales. Accordingly, we have replaced nearly independent from face count with with inference time scaling moderately as face count increases in the revised manuscript. as shown in Line 910 in the revision pdf.

AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

W4 & Q4: Test Set, Checkpoints, Data/Compute Budgets, and Topological Metrics

Thank you for raising these important evaluation questions.

Test-set construction and overlap. Our training corpus consists of approximately 350K meshes. The 500-case evaluation set has zero overlap with the training data: it comprises 200 held-out Objaverse-XL meshes and 100 held-out Toys4K meshes reserved from training, together with 200 out-of-distribution meshes generated by Hunyuan, Tripo, and Direct3D-S2.

Baseline checkpoints and data/compute budgets. Because ARMesh and VertexRegen have not released pretrained weights, we trained both baselines from scratch using their official implementations on approximately 200K meshes (<5K faces) from our dataset, for 5 days each on 8 H800 GPUs (40 GPU-days). For all other baselines, we used their official public checkpoints—preserving their original dataset and compute advantages (Table 4). Notably, most competing baselines were trained on comparable or substantially larger datasets (up to 600K meshes) and significantly higher compute budgets (112–736 GPU-days vs. our 56 GPU-days), confirming that MeshOctave's gains do not stem from larger training data or compute.

Table 4. Training datasets and budgets.

MethodMeshesGPUsDaysGPU-days
MeshAnythingV2≈98K32×A8004128
BPT1.5M→0.3M32×L40≈7224
MeshSilksong≈100K16×H800≈15240
DeepMesh≈500K128×A8004512
MeshRipple≈300K16×A80016256
MeshFlow≈600K32/64×H10013736
LATO.2≈550K8×H10014112
ARMesh≈200K8×H800540
VertexRegen≈200K8×H800540
Ours350K8×H800756

Topological validity and mesh-quality metrics. In Table 5, we evaluate geometric fidelity, topological validity, and mesh quality across all 500 validation meshes, along with average generation time on 50 randomly sampled cases. Because native mesh generation permits open surfaces rather than enforcing watertightness, we report Boundary Edge Ratio (BER) and Non-Manifold Edge Ratio (NMER) alongside the four MeshRFT metrics: Quad Ratio (QR), Angle Quality (AQ), Aspect Quality (ARQ) (which directly penalize degenerate/sliver triangles), and Adjacent Consistency (AC).

MeshOctave achieves the lowest CD and HD and the highest scores across all four mesh-quality metrics. In particular, it substantially outperforms MeshFlow and LATO.2 in QR (0.6175 vs. 0.2840 and 0.2867), AQ (0.8953 vs. 0.8046 and 0.8158), and ARQ (0.8186 vs. 0.7114 and 0.7317), while also achieving higher AC. Moreover, although MeshFlow and LATO.2 exhibit lower BER, they incur 3.5×–4.7× higher NMER (17.10% and 22.89% vs. our 4.85%), demonstrating that MeshOctave achieves a markedly stronger balance of topological cleanliness, triangle regularity, and geometric fidelity.

Table 5. Geometric fidelity, mesh quality, and generation time. Time is in seconds; s/kF denotes seconds per 1,000 output triangles.

MethodCD ↓HD ↓BER↓NMER↓QR ↑AQ ↑ARQ ↑AC ↑Time↓s/kF ↓
VertexRegen0.08740.187516.92251.03790.15940.79660.66130.8385312.53243.53
ARMesh0.08040.15696.87395.68760.34110.83240.72740.7962263.53208.21
BPT0.06100.11112.96410.71630.56870.84710.77180.7998143.0287.32
MeshRipple0.04680.09481.60800.80940.52570.83420.76870.7847298.7395.75
MeshFlow0.04520.07650.536217.10470.28400.80460.71140.78724.585.42
LATO.20.04060.06650.839522.88780.28670.81580.73170.84378.281.40
Nexus0.04170.06555.494122.95720.20090.79610.69690.777213.302.48
Ours0.03920.06451.97024.84840.61750.89530.81860.865360.467.90
Reviewer_Bui4OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MeshOctave, a coarse-to-fine native mesh generation framework based on dyadic spatial grid resolutions. By framing mesh coarsening as deterministic voxel vertex clustering ("resolution collapse"), its inverse ("split-and-rewire") is parameterized per parent face using nine discrete structural tokens: three 8×18\times 1 occupancy vectors, three 8×88\times 8 intra-parent adjacency matrices (SiS_i), and three 8×88\times 8 inter-parent adjacency matrices (CijC_{ij}). An hourglass Transformer trained with a mask-uniform discrete diffusion objective predicts these tokens across resolution scales.

Critical strengths
  1. Unlike two-stage paradigms that decouple vertex generation from connectivity prediction, MeshOctave jointly generates child vertices and local edge topology within a single transition at every scale, yielding valid meshes with explicit topology across intermediate scales and naturally supporting adaptive mesh subdivision.

  2. Defining mesh scales via deterministic dyadic grid quantization rather than sequential QEM-guided edge collapses enables parallel per-face decoding within each scale, bounding scale transitions to log⁡2L\log_2 L and achieving an order-of-magnitude speedup over level-of-detail autoregressive baselines.

Critical weaknesses
  1. Why did the authors choose discrete diffusion over {−1,1}\{-1, 1\} continuous relaxation paradigms like Nexus? While Table 3 includes a Continuous ablation variant, neither the main text nor the appendix explains how this variant is implemented (e.g., target space parameterization, noise schedule, or thresholding). Moreover, running multi-step mask-uniform discrete diffusion across 10 scales results in ~90s inference latency. Could the authors clarify the exact setup of the Continuous ablation and provide a direct quantitative comparison with Nexus?

  2. For any connected parent vertices A–BA\text{--}B, resolving child connectivity only requires predicting small 8×88\times 8 binary matrices (SA,SB,CABS_A, S_B, C_{AB}). Instead of direct element-wise binary prediction, the authors treat each row as a 256-way categorical variable and decode the 8 rows autoregressively (Dc\mathcal{D}_c, Eq. 5). Why is an 8-row 256-way AR decoder preferable here? Furthermore, the intra-parent matrix SiS_i represents an undirected graph among at most 8 children, which is inherently symmetric (Si=SiTS_i = S_i^T) with a zero diagonal (only 28 upper-triangular pairs). Predicting all 64 bits via an 8-row AR head introduces redundancy and does not structurally guarantee symmetry. How are asymmetric entries in SiS_i handled during decoding?

  3. In this cascade formulation, tokens at scale ll are spawned strictly from faces generated at scale l−1l-1. While minor vertex errors are tolerable, edge rewiring errors accumulate severely across scales—once an edge connection is mispredicted at a coarse level, subsequent scales cannot recover from the corrupted topology. Moreover, each parent face decodes its 9 tokens (O0,O1,O2;S0,S1,S2;C01,C02,C12)(O_0, O_1, O_2; S_0, S_1, S_2; C_{01}, C_{02}, C_{12}) locally. Since parent vertices and edges are shared by adjacent faces, how does the method resolve conflicts when neighboring faces predict inconsistent occupancies OiO_i or connectivity matrices (Si,CijS_i, C_{ij}) on shared mesh elements? Given that Table 2 only evaluates surface point cloud metrics (CD, HD, NC), could the authors report quantitative topological metrics—specifically the ratio of broken/unclosed faces, boundary/crack edges, and non-manifold edges—and compare them against the Nexus paradigm?

Questions for authors

See above.

5 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

We sincerely thank the reviewer for the technical assessment and recognition of our joint prediction of vertices and local topology, and order-of-magnitude speedup over level-of-detail autoregressive baselines. Below, we address discrete diffusion, connectivity decoding, cross-scale errors, and topological evaluation.

W1: Choice of Discrete Diffusion, Continuous Ablation Setup, and Comparison with Nexus

Continuous ablation setup. We use an 18-block Transformer with width 768, 12 heads, FFN width 2048. Rectified flow models 8 occupancy components and 64 components per connectivity matrix, with targets in {−1,+1}\{-1,+1\}. Training uses standard Gaussian noise, linear interpolation, and velocity MSE. Inference uses 20 uniform Euler steps over [0,1][0,1], CFG 1.0, and zero-threshold binarization.

Why discrete diffusion over continuous relaxation. We formulate local subdivision as joint categorical pattern prediction rather than independent per-bit regression. An 8-bit occupancy vector has 28=2562^8 = 256 configurations, and each connectivity row likewise has 256 connection patterns. Predicting each complete vector or row as a 256-way category directly captures intra-pattern structural correlations. By contrast, continuous relaxation models entries in continuous space and applies post-hoc thresholding, which can produce structurally incoherent bit combinations and higher cross-face disagreement.

Quantitative results in Table 3 (paper) and qualitative comparisons in Figure 7 (paper) demonstrate the advantage of our discrete formulation. Table 1 reports cross-face inconsistency rates on 200 validation samples, measuring disagreements on shared occupancy vectors (O\mathbf{O}), intra-parent connectivity matrices (S\mathbf{S}), and inter-parent connectivity matrices (C\mathbf{C}) prior to decoding or consensus reconciliation. Continuous incurs higher conflict rates (7.592% / 12.235% / 29.140%) than Mask-Uniform (0.002% / 0.060% / 0.001%). These results support the advantage of joint categorical pattern prediction in preserving structural coherence and reducing cross-face disagreement.

Table 1. Cross-face inconsistency on 200 samples (↓). O: occupancy; S/C: intra-/inter-parent connectivity. Masked Diffusion and Mask-Uniform share a checkpoint and AR head.

MetricContinuousParallel headsMasked DiffusionMask-Uniform
O7.592%2.063%1.221%0.002%
S12.235%23.514%9.819%0.060%
C29.140%10.839%3.451%0.001%

Comparison with Nexus and inference latency. During method development, we faithfully reproduced Nexus following its paper and trained it on the same 350K-mesh dataset using 8 GPUs for 7 days. Table 2 reports geometric fidelity, topological validity, and mesh quality across the full 500-case validation set, along with inference speed averaged over 50 randomly sampled validation cases. Compared to Nexus, MeshOctave achieves superior geometric fidelity (CD: 0.0392 vs. 0.0417; HD: 0.0645 vs. 0.0655), substantially fewer boundary and non-manifold defects (BER: 1.97% vs. 5.49%; NMER: 4.85% vs. 22.96%), and markedly higher mesh quality across all four MeshRFT metrics (e.g., QR: 0.6175 vs. 0.2009). This gain comes with higher runtime than Nexus (60.46 s vs. 13.30 s per mesh; 7.90 vs. 2.48 s/kF), reflecting a trade-off between raw sampling speed and structural/topological fidelity—while remaining 2.4×, 5.2× faster per mesh than autoregressive and next-scale baselines (BPT, ARMesh, MeshRipple, VertexRegen). Revised Fig. 11 further extends our inference-time scaling comparison across output face counts to all baselines.

Table 2. Fidelity, topology, quality, and time. Boundary edge ratio (BER) and non-manifold edge ratio (NMER) are percentages. Quad ratio (QR), angle quality (AQ), aspect ratio quality (ARQ), and adjacent consistency (AC) follow Mesh-RFT. Time: seconds; s/kF: seconds per 1,000 triangles.

MethodCD ↓HD ↓BER↓NMER↓QR ↑AQ ↑ARQ ↑AC ↑Time↓s/kF ↓
VertexRegen0.08740.187516.92251.03790.15940.79660.66130.8385312.53243.53
ARMesh0.08040.15696.87395.68760.34110.83240.72740.7962263.53208.21
BPT0.06100.11112.96410.71630.56870.84710.77180.7998143.0287.32
MeshRipple0.04680.09481.60800.80940.52570.83420.76870.7847298.7395.75
MeshFlow0.04520.07650.536217.10470.28400.80460.71140.78724.585.42
LATO.20.04060.06650.839522.88780.28670.81580.73170.84378.281.40
Nexus0.04170.06555.494122.95720.20090.79610.69690.777213.302.48
Ours0.03920.06451.97024.84840.61750.89530.81860.865360.467.90
AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

W2: AR connectivity head, symmetry, and redundancy

Thank you for these constructive questions.

Why an 8-row 256-way AR head instead of element-wise binary prediction. Our design is motivated by categorical token prediction in autoregressive models: given a 256-entry vocabulary, predicting a single categorical index selects a coherent 8-bit connection pattern jointly rather than assuming conditional independence across bits. Decoding the 8 rows autoregressively then allows each row to condition on previously predicted rows, coordinating the global structure of the 8×88 \times 8 adjacency matrix. Predicting the entire matrix in one step would require an intractable 2642^{64}-way vocabulary, making row-wise 256-way AR decoding a practical balance between structured joint prediction and tractable classification.

Importantly, the 8 sequential steps occur only within the lightweight MLP output head; all matrices within a scale are processed in parallel without re-running the Transformer backbone. The necessity of row-wise autoregression is further confirmed by the Parallel Heads ablation in Table 1 (where the 8 rows are predicted independently in parallel): removing row-wise conditioning causes cross-face inconsistency on intra-parent connectivity S\mathbf{S} to jump from 0.060% to 23.514% (and on C\mathbf{C} from 0.001% to 10.839%).

Symmetry and redundancy in S\mathbf{S}. The reviewer is completely right that S\mathbf{S} represents an undirected graph with zero diagonal and only 28 upper-triangular entries. Currently, our 8-row AR head does not structurally restrict the output parameterization to these 28 entries, exploring a minimal upper-triangular parameterization is an interesting avenue for future optimization. During decoding, we explicitly enforce symmetry via element-wise logical OR, S←S∨S⊤\mathbf{S} \leftarrow \mathbf{S} \lor \mathbf{S}^{\top}. In practice, however, the AR decoder virtually rarelly violates symmetry: as reported in Table 3, across 200 validation samples, 100.00% of all 651,630 raw predicted S\mathbf{S} matrices were already strictly symmetric with zero diagonals before occupancy masking, symmetrization, or decoding.

Table 3. Raw S prediction statistics.

MatricesSym. (%)Asym. (%)Zero diag. (%)
651,630100.000.00100.00
AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

W3: Cross-scale errors, shared conflicts, and topology

Thank you for raising these important concerns.

Cross-scale error accumulation. MeshOctave predicts 10 hierarchical scales sequentially, and as discussed in Sec. 6, coarse-scale errors can occasionally propagate to finer scales. However, several architectural properties limit this effect:

  1. Short cascade depth: A 10-scale hierarchy has a dramatically shorter sequential dependency chain than autoregressive mesh generation methods (e.g., ARMesh and VertexRegen), which must sequentially predict hundreds or thousands of individual local refinements.
  2. Intra-scale revision: Within each scale, the mask pathway proposes initial structures, the uniform pathway revises predictions, and confidence-based re-masking revisits uncertain tokens before advancing to the next scale.
  3. Scaling behavior: Consistent with trends observed in large autoregressive models, we empirically observe that increasing model capacity and training data substantially reduces coarse-scale error propagation.

Resolving shared vertex and connectivity conflicts. In Table 1, we measure the proportion of shared mesh elements exhibiting any disagreement across adjacent parent faces (comparing O\mathbf{O} and S\mathbf{S} across incident faces sharing a parent vertex, and C\mathbf{C} across faces sharing a parent edge), rather than per-entry mismatch rates. Under Mask-Uniform diffusion, these element-level conflict rates are only 0.002% (O\mathbf{O}), 0.060% (S\mathbf{S}), and 0.001% (C\mathbf{C}). After the final denoising step at each scale, any remaining disagreements are resolved via majority voting over complete vectors/matrices (selecting the most frequent candidate, with deterministic tie-breaking). For inter-parent connectivity C\mathbf{C}, matrices are first aligned to a canonical parent-edge orientation before voting and transposed back afterward. While this enforces exact cross-face agreement on shared structural tokens, we note that it does not strictly guarantee global manifoldness.

Topological metrics vs. Nexus. Because MeshOctave targets native mesh generation and permits open surfaces rather than enforcing watertightness, we evaluate topological defects using Boundary Edge Ratio (BER) and Non-Manifold Edge Ratio (NMER) alongside four MeshRFT mesh-quality metrics in Table 4. Compared to Nexus, MeshOctave reduces BER by nearly 2.8× (1.97% vs. 5.49%) and NMER by over 4.7× (4.85% vs. 22.96%). Furthermore, while continuous flow baselines such as MeshFlow and LATO.2 exhibit lower BER, they incur 3.5×–4.7× higher NMER than MeshOctave (17.10% and 22.89% vs. 4.85%), demonstrating that MeshOctave achieves a substantially stronger balance of topological cleanliness, geometric fidelity, and mesh quality.

Table 4. Geometric fidelity, topology, and mesh quality. Boundary edge ratio (BER) and non-manifold edge ratio (NMER) are percentages. Quad ratio (QR), angle quality (AQ), aspect quality (ARQ), and adjacent consistency (AC) follow Mesh-RFT.

MethodCD ↓HD ↓BER↓NMER↓QR ↑AQ ↑ARQ ↑AC ↑
VertexRegen0.08740.187516.92251.03790.15940.79660.66130.8385
ARMesh0.08040.15696.87395.68760.34110.83240.72740.7962
BPT0.06100.11112.96410.71630.56870.84710.77180.7998
MeshRipple0.04680.09481.60800.80940.52570.83420.76870.7847
MeshFlow0.04520.07650.536217.10470.28400.80460.71140.7872
LATO.20.04060.06650.839522.88780.28670.81580.73170.8437
Nexus0.04170.06555.494122.95720.20090.79610.69690.7772
Ours0.03920.06451.97024.84840.61750.89530.81860.8653
Reviewer_a46UOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MeshOcTave, a vertex splitting method to construct a high resolution mesh from a point cloud.  The method make use of a multi-resolution grid representation for producing the training data.  The high resolution grid structure is prepared that includes the original high resolution mesh. The resolution of the mesh is gradually decreased.  When dropping the resolution, all the vertices in the same grid are merged into a single vertex, and the mesh connectivity is updated accordingly.  This process is continued until the grid becomes a single cell.  The inverse process of constructing the higher resolution mesh from the lower resolution mesh is learned by a diffusion model.  The point cloud of the object is given as a condition for this diffusion model.   The lower resolution mesh can be also used as a mesh to start the vertex splitting.  The system is trained on a large dataset of 350k meshes.   The evaluation is done by comparing the geometry of the synthesized mesh with the ground truth data.  The system is compared with the other baselines such as MeshFlow and Lato2, using criteria such as CD-L1・CD-L2・Hausdorff distance・Normal consistency - and the proposed method shows better numbers with all the criteria.

Critical strengths
  • A novel diffusion-based mesh synthesis method based on multi-resolution grid.
  • Showing the best performance compared to the other baselines on the four criteria.
  • Faster inference time compared to the baselines.
Critical weaknesses
  1. Topological Validity: The authors should report manifoldness and watertightness metrics, as geometric similarity metrics alone do not establish the topological validity of generated meshes.

  2. Connectivity Consistency: How does the method ensure global topological consistency when independently predicted connectivity structures for adjacent faces are assembled into a complete mesh?

Questions for authors
  • The way to compute the vertex location in the low resolution from the vertices of the higher resolution mesh is not described - it is only described they are merged but whether it is the average of the vertices or the center of the grid etc. is not described.  - The comparison to the baseline for inference is done with ARMesh / VertexRegen but not with MeshFLow and Lato2- why is that?
4 comments
AuthorsOpenReview ↗

Rebuttal by Authors [1/3]

We sincerely thank the reviewer for the thoughtful and constructive feedback, as well as for recognizing our method's novelty, superior performance, and faster inference speed relative to the compared baselines. Below, we address each of your concerns regarding topological validity, connectivity consistency, and methodological clarity point by point.

W1: Topological validity - manifoldness, watertightness

Thank you for raising this important point. We would like to clarify that MeshOctave targets native (artist-style) mesh generation, and therefore naturally accommodates non-manifold edges and open surfaces, consistent with both autoregressive baselines (e.g., MeshAnything V2, BPT, ARMesh, VertexRegen, DeepMesh, MeshRipple) and flow-based native mesh methods (e.g., MeshFlow, LATO.2). Relaxing the strict watertightness constraints, typically imposed by implicit or field-based reconstruction approaches, enables our model to train directly on vast, general-purpose mesh collections rather than being restricted to watertight assets.

To evaluate topological validity and mesh quality, we additionally report Boundary Edge Ratio (BER) and Non-Manifold Edge Ratio (NMER), alongside four mesh-quality metrics from Mesh-RFT: Quad Ratio (QR), Angle Quality (AQ), Aspect Ratio Quality (ARQ), and Adjacent Consistency (AC). We evaluate all methods on the full 500-case validation set and report the results in Table 1, alongside Chamfer Distance (CD) and Hausdorff Distance (HD) for completeness. Notice, manifoldness and watertightness metrics may yield unfavorable assessments for native mesh generators, as they penalize valid open boundaries and non-manifold artist structures.

Topology and mesh quality. MeshOctave achieves the lowest CD and HD and the highest scores across all four mesh-quality metrics among all compared methods. In particular, it substantially outperforms MeshFlow and LATO.2 in QR (0.6175 vs. 0.2840 and 0.2867), AQ (0.8953 vs. 0.8046 and 0.8158), and ARQ (0.8186 vs. 0.7114 and 0.7317), while also achieving higher AC. Moreover, although MeshFlow and LATO.2 exhibit lower BER, they incur 3.5×–4.7× higher NMER (17.10% and 22.89% vs. our 4.85%). These results demonstrate that MeshOctave improves mesh quality while maintaining strong geometric fidelity.

Table 1. Geometric fidelity and mesh quality.

MethodCD ↓HD ↓BER ↓NMER ↓QR ↑AQ ↑ARQ ↑AC ↑
VertexRegen0.08740.187516.92251.03790.15940.79660.66130.8385
ARMesh0.08040.15696.87395.68760.34110.83240.72740.7962
BPT0.06100.11112.96410.71630.56870.84710.77180.7998
MeshRipple0.04680.09481.60800.80940.52570.83420.76870.7847
MeshFlow0.04520.07650.536217.10470.28400.80460.71140.7872
LATO.20.04060.06650.839522.88780.28670.81580.73170.8437
Nexus0.04170.06555.494122.95720.20090.79610.69690.7772
Ours0.03920.06451.97024.84840.61750.89530.81860.8653
AuthorsOpenReview ↗

Rebuttal by Authors [2/3]

W2: Consistency across adjacent faces

Thank you for this insightful question. In MeshOctave, each parent-face block is decoded locally, and coincident child vertices are merged via quantization to grid coordinates. Crucially, our Mask-Uniform discrete diffusion process mitigates cross-face discrepancies during generation, and ensure the mesh completeness: because it permits revising previously unmasked tokens within each scale, it detects conflicting local predictions across neighboring faces through confidence scores and iteratively resolves them through re-masking and denoising.

To validate this mechanism empirically, we measure cross-face disagreement rates on 200 randomly selected validation samples. Specifically, we compare occupancy OO and intra-parent connectivity SS predictions for the same parent vertex across all incident parent faces, and we compare inter-parent connectivity CC predictions for child vertices along the endpoints of each shared parent edge.

As shown in Table 2, replacing standard Masked Diffusion with Mask-Uniform diffusion dramatically reduces cross-face inconsistencies across all three token types. After the final denoising step at each scale, any remaining disagreements are deterministically reconciled by assigning all shared occurrences the most frequent complete occupancy vector or connectivity matrix, with deterministic tie-breaking (for inter-parent matrices, predictions are first aligned to a canonical parent-edge orientation and transposed back afterward). This post-denoising reconciliation guarantees exact structural agreement across shared elements.

Table 2. Cross-face inconsistency ratio (↓).

MetricMasked DiffusionMask-Uniform
Occupancy (O\mathbf{O})1.221%0.0019%
Intra-parent (S\mathbf{S})9.819%0.0596%
Inter-parent (C\mathbf{C})3.451%0.0011%
AuthorsOpenReview ↗

Rebuttal by Authors [3/3]

Q1: Low-resolution vertex placement and inference comparisons

Thank you for pointing out these areas, where further clarification is helpful.

Low-resolution vertex placement. Vertices are placed at occupied cell centers rather than averages of merged vertex positions. After normalization to the cube [-1,1]^3, the center of a cell with integer grid coordinates q_l at resolution R = 2^l is given coordinate-wise by: x_l = 2 * (q_l + 0.5) / R - 1.

During resolution collapse, child cells map to parent cells via q_(l-1) = floor(q_l / 2), and vertices mapped to the same parent cell are merged. During generation, the eight candidate child cells have coordinates 2 * q_(l-1) + delta, where delta ranges over {0,1}^3. Occupancy predictions select which children are instantiated, with each child vertex placed at its cell center. We have clarified this placement rule in Section 3.2 of the revised manuscript.

Inference efficiency. We focused our runtime comparison on ARMesh and VertexRegen because they share our next-scale native mesh generation paradigm and thus serve as our most direct architectural baselines. Following your suggestion, we have expanded our runtime evaluation in Table 3 to include MeshFlow, LATO.2, Nexus, and additional autoregressive baselines (BPT and MeshRipple), reporting both average wall-clock time per mesh and seconds per 1,000 output triangles (s/kF) over 50 randomly sampled validation cases.

As shown in Table 3, MeshOctave is substantially faster than all autoregressive and next-scale baselines,e.g., achieving a 4.4×–5.2× speedup per mesh over ARMesh and VertexRegen and a 2.4×–4.9× speedup over BPT and MeshRipple (with an order-of-magnitude advantage in s/kF), while remaining slower than single-step/continuous flow methods such as MeshFlow and LATO.2. This reflects a deliberate trade-off between raw inference speed and superior geometric accuracy and topological quality (reflected by a composite topology score TS of 75.82 vs. 56.67 and 58.17). In addition, Revised Fig. 11 extends our scaling analysis of inference time versus output face count, originally limited to ARMesh and VertexRegen, across all baselines.

Table 3. Geometric quality, mesh quality, and inference time. CD, HD, and the composite mesh-quality score TS are evaluated on all 500 validation samples. TS = 100 × (0.4 QR + 0.2 AQ + 0.3 ARQ + 0.1 AC). Time is averaged over 50 randomly selected cases. s/kF denotes seconds per 1,000 output triangles.

MethodCD ↓HD ↓TS ↑Time (s) ↓s/kF ↓
ARMesh0.08040.156960.08263.53208.21
VertexRegen0.08740.187550.53312.53243.53
BPT0.06100.111170.84143.0287.32
MeshRipple0.04680.094868.62298.7395.75
MeshFlow0.04520.076556.674.585.42
LATO.20.04060.066558.178.281.40
Nexus0.04170.065552.6413.302.48
Ours0.03920.064575.8260.467.90
Reviewer_a46UOpenReview ↗

concerns are addressed

Thank you very much for the rebuttal. My concerns are addressed with these comments.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Authors’ Summary of Revisions and Responses to Reviewer Concerns

We sincerely thank all reviewers for their thoughtful and constructive feedback. Below, we summarize the major manuscript revisions. All line numbers refer to the revised PDF; W and Q denote reviewer weaknesses and questions. All newly added or modified text is highlighted in blue.

1. Cross-Face Conflict Rates and Deterministic Reconciliation

  • Section 4.1, “Inference” (lines 340–343); Section 5.3 (Table 5, lines 479–485); Appendix A.9 (lines 1021–1023) — Reviewer a46U, W2; Reviewer Bui4, W1–W3; Reviewer YeSi, W1 & Q1: We added cross-face inconsistency measurements on 200 samples, showing conflict rates of only 0.002%, 0.060%, and 0.001% for occupancy, intra-parent connectivity, and inter-parent connectivity, respectively. We explicitly described the deterministic reconciliation procedure: after final denoising at each scale, conflicting shared predictions are resolved by majority voting over complete vectors or matrices with deterministic tie-breaking. Inter-parent matrices are aligned to a canonical parent-edge orientation before voting and restored afterward.

2. Expanded Topological, Mesh-Quality, and Efficiency Evaluation

  • Section 5.1 (Table 3, lines 395–406); Appendix A.8 (Table 7, lines 969–1001) — Reviewer a46U, W1; Reviewer Bui4, W1 & W3; Reviewer YeSi, W4 & Q4: We expanded evaluation beyond geometric fidelity to include boundary edge ratio (BER), non-manifold edge ratio (NMER), and four Mesh-RFT mesh-quality metrics: QR, AQ, ARQ, and AC. The extended comparison evaluates quality on all 500 cases and includes our Nexus reproduction. MeshOctave achieves the highest scores across all four mesh-quality metrics and lower BER and NMER than Nexus. We also clarified that native mesh generation permits open surfaces and non-manifold assets, and that shared-token agreement does not imply global manifoldness or watertightness.

  • Section 5.1, “Inference Latency” (lines 407–410); Appendix A.7 (lines 907–948; Figure 11, lines 960–966); Appendix A.8 (Table 7) — Reviewer a46U, Q1; Reviewer Bui4, W1; Reviewer YeSi, W3 & Q3: We broadened inference comparisons to additional autoregressive and flow-based baselines, reporting average generation time and seconds per 1,000 output triangles, together with runtime and peak allocated GPU memory versus output face count. We specified the measurement hardware, batch size, and timing scope, and identified the ARMesh and VertexRegen resource curves as estimates. We replaced the imprecise face-count independence claim with “scales moderately with face count” and explicitly discussed the speed–quality trade-off relative to faster flow-based methods and Nexus.

3. Methodological and Experimental Clarifications

  • Section 3.2 (lines 210–214 and 254–257); Section 5.3 (lines 466–474); Appendices A.10–A.11 (lines 1058–1107) — Reviewer a46U, Q1; Reviewer Bui4, W1 & W2; Reviewer YeSi, W1 & Q1: We added explicit child-vertex coordinates and the triangle-construction rule, completed the continuous-ablation training and inference settings, and explained the choice of discrete categorical modeling over continuous relaxation. We also clarified why row-wise categorical autoregression is used instead of independent binary heads, emphasizing joint row-pattern modeling and cross-row dependencies. The appendix further specifies symmetry enforcement and diagonal clearing for intra-parent connectivity, supported by statistics on 651,630 raw predicted matrices.

  • Appendix A.9 (lines 1005–1054; Tables 8–9) — Reviewer YeSi, W2 & Q2 and W4 & Q4: We clarified the disjoint 500-case evaluation split, baseline checkpoint sources, training data sizes, hardware, and training durations. We documented the Nexus reproduction on the same 350K-mesh dataset using eight H800 GPUs for seven days. We also explicitly distinguished the 624M-parameter main backbone from the 276M-parameter ablation backbone, explaining the different “Full” results and training budgets. Public-checkpoint comparisons retain their original training settings, and GPU-days are reported without hardware normalization.

Dear Reviewers,

A friendly reminder to check the authors’ rebuttals and engage in the discussion, as the deadline is approaching. Authors still have time to address any further questions or concerns you may have, so your timely feedback would be greatly appreciated.

Thank you for your time and effort!

Your AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.