MeshOctave generates meshes via cascading resolution transitions

Junkai Lin; Tianhao Zhao; Hang Long; Huipeng Guo; Jielei Zhang; Youjia Zhang; Jiale Xu; Wenbing Li

cs.CVSep 30, 2026

MeshOctave generates meshes via cascading resolution transitions

Authors: Junkai Lin, Tianhao Zhao, Hang Long, Huipeng Guo, Jielei Zhang, Youjia Zhang, Jiale Xu, Wenbing Li, +5 more

Organizations: Huazhong University of Science and Technology · Meshy AI · Independent Researcher · Technical University of Munich

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.

Figures & tables

Figure 1: MeshOctave generates meshes via cascading resolution transitions. Top: Starting from a single voxel, each step doubles vertex resolution by predicting child vertex instantiation ( split ) and reconnection ( rewire ) for all parent faces in parallel. Bottom: Meshes generated by MeshOctave.

Panel of Table 1 — Table 1: Comparison with prior works. MeshOctave jointly refines vertices and topology via multi-scale raw-space masked diffusion. Notation: Rep. R : raw; L : latent; Method A : autoregressive; D : diffusion/flow; M : masked diffusion; Arch. N : next scale; W : whole mesh.

Figure 2: Face-centric split-and-rewire. Resolution collapse merges vertices fall in the same grid cell; its inverse, split-and-rewire specifies which of the eight child voxels are occupied and how the instantiated children are connected. For each parent face, 3 8-bit occupancy vectors encode vertex splitting, and 3 intra-parent and 3 inter-parent $8\times 8$ connectivity matrices encode rewiring.

Figure 3: Scale-conditioned discrete diffusion model. The generator refines a mesh from scale $l$ to $l+1$ , the split-and-rewire structural tokens $\mathbf{z}$ is concatenated with vertex coordinates. The elements of $\mathbf{z}$ are masked according to diffusion timestep $t$ . An hourglass Transformer processes these tokens, modulated by timestep $t$ and scale $l$ via adaLN. Finally, categorical heads predict child-vertex occupancy while an autoregressive MLP head decodes row-factorized adjacency matrices, assembling the refined mesh at scale $l+1$ .

Metric	Mesh Any.V2	Vertex Regen	AR Mesh	Fast Mesh	Mesh Silk.	BPT	Deep Mesh	Mesh Ripple	Mesh Flow	LATO	LATO.2	MESH OCTAVE
Type	AR	AR+ N	AR+ N	AR	AR	AR	AR	AR	Flow	Flow	Flow	DD+ N
CD-L2 $\downarrow$	0.1103	0.0874	0.0804	0.0647	0.0622	0.0610	0.0505	0.0468	0.0452	0.0430	0.0406	0.0392
CD-L1 $\downarrow$	0.1539	0.1238	0.1146	0.0925	0.0895	0.0873	0.0729	0.0691	0.0672	0.0621	0.0595	0.0581
HD $\downarrow$	0.2317	0.1875	0.1569	0.1151	0.1440	0.1111	0.0957	0.0948	0.0765	0.0741	0.0665	0.0645
$\|\mathrm{NC}\|$ $\uparrow$	0.6807	0.6655	0.6058	0.7413	0.7728	0.7976	0.8187	0.8141	0.8237	0.8273	0.8333	0.8478

Table 2: Quantitative comparison. Type denotes the generative paradigm: autoregressive generation (AR), continuous flow matching (Flow), or discrete diffusion (DD). N marks methods that generate meshes over a coarse-to-fine resolution hierarchy and expose intermediate scales. Best results are in bold , and second-best results are underlined .

Figure 4: Qualitative comparison. Each row shows one input point cloud and the final meshes generated by autoregressive, flow-matching, and diffusion-based methods.

Figure 5: Comparison with level-of-detail baselines.

Figure 6: Generation process visualization. Meshes are progressively generated across scales, while predictions within a scale are iteratively refined through masked and uniform denoising.

Figure 8: Application in mesh subdivision. MeshOctave adaptively refines geometry and connectivity given an existing coarse mesh, producing more detailed and artist-like subdivision results.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Hourglass level	Token Corresponds to	Anchor	Type $\tau$
$9/$ face, coordinate-level	$F_{0},F_{1},F_{2}$	$(v_{0},v_{0}),(v_{1},v_{1}),(v_{2},v_{2})$	$0$
	$M^{\mathrm{intra}}_{0},M^{\mathrm{intra}}_{1},M^{\mathrm{intra}}_{2}$	$(v_{0},v_{0}),(v_{1},v_{1}),(v_{2},v_{2})$	$1$
	$M^{\mathrm{inter}}_{01},M^{\mathrm{inter}}_{02},M^{\mathrm{inter}}_{12}$	$(v_{0},v_{1}),(v_{0},v_{2}),(v_{1},v_{2})$	$2$
$3/$ face, vertex-level	subdivision / intra / inter aggregate	$(v_{0},v_{1},v_{2})$	$0/1/2$
$1/$ face, face-level	face feature	$(v_{0},v_{1},v_{2})$	$3$

Appendix

Table 4: Geometry-anchored rotary encoding. The

30

geometric slots are split as

2

anchors

\times\ 3

axes

\times\ 5

slots at

9/

face, and as

3

vertices

\times\ 3

axes

\times\ 3

slots after reduction.

\tau

separates tokens that share an anchor.

Figure 9: Intra-scale denoising. The left column shows the resolution-32 parent mesh. The remaining columns show the decoded predictions immediately before re-masking at diffusion steps 0, 4, 8, 14, and 20 of the same resolution transition.

Figure 10: Mesh generation across spatial resolutions. Intermediate meshes from one coarse-to-fine generation trajectory. Each column represents the completed mesh at a different grid resolution. Increasing resolution progressively adds geometric detail and refines connectivity; the result at each level can be used as a standalone level-of-detail mesh.

Figure 11: Inference wall-clock time versus output mesh complexity (number of faces), for the next-scale baseline (armesh), the autoregressive vertex-split baseline (VertexRegen), and MeshOctave . MeshOctave generates comparably complex meshes over an order of magnitude faster than both baselines. All curves are directly measured. Neither baseline was trained beyond

5\text{k}

faces (dotted vertical line); their dashed segments are still measured wall-clock time, but the decoded mesh’s validity at that scale is unverified. The MeshOctave curve is solid throughout: every bin is a directly measured, verified generation.

Explore similar work

PolyFlow: Continuous Topology Embedding Flow Matching for Artist-style Mesh Generation
Jun 25, 2026Chunshi Wang, Haohan Weng, Junliang Ye +9Mesh Generation Riemannian Flow Matching
1ZJU · 2Tencent · 3THU +3
Nexus: Native Mesh Generation with Diffusion
Jul 15, 2026Hanxiao Wang, Ying-Tian Liu, Yuan-Chen Guo +5Mesh Generation Voxel
MAIS, Institute of Automation, Chinese Academy of Sciences; School of Artificial Intelligence, University of Chinese Academy of Sciences, China · VAST, China · Tsinghua University, China +1
MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation
Jun 3, 2026Jiale Xu, Wang Zhao, Ying ShanMesh Generation Voxel
ARC Lab, Tencent PCG

Jun 25, 2026, cs.GRJ/K move · Enter open · S save

PolyFlow: Continuous Topology Embedding Flow Matching for Artist-style Mesh Generation

Chunshi Wang, Haohan Weng, Junliang Ye, Biwen Lei+8

1ZJU · 2Tencent · 3THU+3

Mesh Generation Riemannian Flow Matching Generative Models Iterative Denoising Methods

Open paper AskPDF arXiv

Autoregressive Transformers dominate high-quality mesh generation by producing artist-worthy topologies, yet their inherent sequential decoding induces substantial computational overhead, falling orders of magnitude slower than parallel generative models. On the other hand, while continuous diffusion and flow-matching methods support efficient parallel synthesis across a variety of domains, they cannot be directly applied to meshes: mesh connectivity is inherently discrete and incompatible with standard continuous noise injection and denoising operations. To resolve this fundamental incompatibility, we introduce a compact topology embedder that projects discrete mesh vertex positions and normals into continuous per-vertex embeddings, where the original discrete adjacency information can be faithfully recovered via spacetime distance thresholding. After pretraining and freezing this embedder, any raw mesh can be fully converted into a continuous per-vertex state space unifying position, normal, and implicit topological attributes. Built upon this novel continuous mesh representation, we present PolyFlow, a Transformer-based flow-matching framework that achieves fully parallel vertex state denoising conditioned on extracted point-cloud features. During inference, our model completes generation rapidly via an ODE solver, and supports explicit, precise control over output mesh resolution by directly specifying the target vertex count. Extensive evaluations on the Toys4K benchmark demonstrate that PolyFlow surpasses state-of-the-art autoregressive baselines in both Chamfer Distance and Hausdorff Distance.

Abstract

Explore similar work

PolyFlow: Continuous Topology Embedding Flow Matching for Artist-style Mesh Generation

Nexus: Native Mesh Generation with Diffusion

MeshWeaver: Sparse-Voxel-Guided Surface Weaving for Autoregressive Mesh Generation