- 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
-
(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?
-
(Relates to weakness 2) Could the authors explain the number gap between the Table 2 and Table 3 “Full” results?
-
(Relates to weakness 3) Could the authors report wall-clock time and memory against all possible baselines as a function of face count?
-
(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
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.
| Method | 0–2k faces | 2–4k faces | 4–9.5k faces |
|---|---|---|---|
| MeshAnythingV2 | 52.1 / 3.81 | 102.3 / 4.72 | — |
| BPT | 101.1 / 2.07 | 180.0 / 2.58 | — |
| DeepMesh | 29.3 / 5.34 | 167.3 / 5.73 | 280.2 / 5.73 |
| Mesh-Silksong | 43.9 / 1.68 | 195.3 / 2.59 | 306.7 / 3.23 |
| MeshRipple | 107.9 / 4.62 | 295.5 / 4.63 | 729.0 / 4.65 |
| MeshFlow | 4.4 / 6.57 | 5.6 / 6.57 | — |
| LATO.2 | 7.2 / 5.24 | — | 8.4 / 5.48 |
| ARMESH | 193.0 / 3.05 | 583.5 / 3.82 | — |
| VertexRegen | 372.3 / 4.24 | 664.5 / 5.09 | — |
| Ours | 30.3 / 2.39 | 45.7 / 3.10 | 84.8 / 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.
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.
| Method | Meshes | GPUs | Days | GPU-days |
|---|---|---|---|---|
| MeshAnythingV2 | ≈98K | 32×A800 | 4 | 128 |
| BPT | 1.5M→0.3M | 32×L40 | ≈7 | 224 |
| MeshSilksong | ≈100K | 16×H800 | ≈15 | 240 |
| DeepMesh | ≈500K | 128×A800 | 4 | 512 |
| MeshRipple | ≈300K | 16×A800 | 16 | 256 |
| MeshFlow | ≈600K | 32/64×H100 | 13 | 736 |
| LATO.2 | ≈550K | 8×H100 | 14 | 112 |
| ARMesh | ≈200K | 8×H800 | 5 | 40 |
| VertexRegen | ≈200K | 8×H800 | 5 | 40 |
| Ours | 350K | 8×H800 | 7 | 56 |
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.
| Method | CD ↓ | HD ↓ | BER↓ | NMER↓ | QR ↑ | AQ ↑ | ARQ ↑ | AC ↑ | Time↓ | s/kF ↓ |
|---|---|---|---|---|---|---|---|---|---|---|
| VertexRegen | 0.0874 | 0.1875 | 16.9225 | 1.0379 | 0.1594 | 0.7966 | 0.6613 | 0.8385 | 312.53 | 243.53 |
| ARMesh | 0.0804 | 0.1569 | 6.8739 | 5.6876 | 0.3411 | 0.8324 | 0.7274 | 0.7962 | 263.53 | 208.21 |
| BPT | 0.0610 | 0.1111 | 2.9641 | 0.7163 | 0.5687 | 0.8471 | 0.7718 | 0.7998 | 143.02 | 87.32 |
| MeshRipple | 0.0468 | 0.0948 | 1.6080 | 0.8094 | 0.5257 | 0.8342 | 0.7687 | 0.7847 | 298.73 | 95.75 |
| MeshFlow | 0.0452 | 0.0765 | 0.5362 | 17.1047 | 0.2840 | 0.8046 | 0.7114 | 0.7872 | 4.58 | 5.42 |
| LATO.2 | 0.0406 | 0.0665 | 0.8395 | 22.8878 | 0.2867 | 0.8158 | 0.7317 | 0.8437 | 8.28 | 1.40 |
| Nexus | 0.0417 | 0.0655 | 5.4941 | 22.9572 | 0.2009 | 0.7961 | 0.6969 | 0.7772 | 13.30 | 2.48 |
| Ours | 0.0392 | 0.0645 | 1.9702 | 4.8484 | 0.6175 | 0.8953 | 0.8186 | 0.8653 | 60.46 | 7.90 |
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 vectors, S matrices, and 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, 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.
Face construction and manifoldness. Each parent triangle has at most 24 candidate child vertices. After occupancy masking, we assemble the S and C blocks into a local undirected adjacency matrix A. For each triple of occupied vertices i<j<k, we emit a triangle if and only if all three edges are present, that is, AijAjkAik=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).