Neural Mesh Models: Analysis and Applications

• Neural Mesh Models are neural architectures that treat mesh structures as primary objects, preserving connectivity and geometric detail for diverse 3D applications.
• They encompass paradigms such as fixed-connectivity deformation, feature-field augmentation, and autoregressive sequence modeling to optimize mesh quality and performance.
• These models enable robust 3D pose estimation, generative design, CFD surrogate modeling, and even neural data analysis, highlighting their broad technical impact.

Searching arXiv for recent and foundational papers on neural mesh models and closely related mesh-native neural representations. Search query: "all:neural mesh models OR ti:mesh draping OR ti:MeshXL OR ti:NeMo neural mesh OR ti:SpaceMesh OR ti:MeshMamba" Neural mesh models are neural architectures in which meshes are treated as primary computational objects rather than as incidental outputs of iso-surfacing, remeshing, or post hoc reconstruction. In the literature, the term does not denote a single canonical model family. It spans explicit neural deformation fields acting on fixed mesh connectivity, category-level generative models with neural features attached to mesh vertices, autoregressive and state-space models over serialized mesh coordinates, graph and equivariant networks operating directly on mesh connectivity, direct manifold-connectivity generators, and volumetric tetrahedral mesh generators for simulation-ready outputs (Hertz et al., 2021, Wang et al., 2021, Chen et al., 2024, Shen et al., 2024, Zheng et al., 2022). Earlier uses also appeared in neural data analysis and recurrent architectures, which indicates that “neural mesh” has historically been a broad label rather than a uniquely settled technical term (Ozay et al., 2012, Onal et al., 2016, Beck et al., 2018).

1. Historical scope and terminological breadth

Some of the earliest “mesh” models in the cited corpus are not graphics models at all. “Mesh Learning for Classifying Cognitive Processes” constructs a star mesh around each fMRI voxel, estimates arc weights from local voxel intensity relations, and uses Mesh Arc Descriptors as classifier inputs for k-NN, SVM, Gaussian Naïve Bayes, and Neural Networks (Ozay et al., 2012). “Modeling the Sequence of Brain Volumes by Local Mesh Models for Brain Decoding” extends this line by forming spatially and functionally local meshes around each seed voxel, fitting ridge-regression weights from neighboring voxel time series, and using the resulting edge weights as features for SVM-based decoding of cognitive tasks (Onal et al., 2016). In both cases, the “mesh” is a local graph over neural measurements rather than a polygonal surface.

A distinct usage appears in “Neural Mesh: Introducing a Notion of Space and Conservation of Energy to Neural Networks,” where the hidden state is a $2$D toroidal grid of neurons, each neuron connects only to its four neighbors, and activation transfer is constrained by a conservation-of-energy–like rule (Beck et al., 2018). Here the mesh is a recurrent computational substrate with spatial locality and persistent excitation, not a geometric surface.

The modern 3D-geometry usage becomes explicit in papers such as “Mesh Draping: Parametrization-Free Neural Mesh Transfer,” where a neural mesh model is a deformation function

$$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$

applied to source mesh vertices while keeping connectivity fixed (Hertz et al., 2021). In “NeMo: Neural Mesh Models of Contrastive Features for Robust 3D Pose Estimation,” the term denotes a category-level mesh $\Gamma_y$ augmented with learned feature vectors $\Theta_y$ at vertices, rendered differentiably and matched to image features (Wang et al., 2021). This later usage has become the dominant one in 3D vision and geometry processing, but the broader record shows that the phrase “neural mesh model” remains intrinsically polysemous.

2. Representational paradigms

A useful way to organize the literature is by the object that is learned or generated: deformations of an existing mesh, neural features attached to a mesh scaffold, tokenized coordinate sequences, graph-native mesh operators, continuous connectivity spaces, implicit dynamic surfaces, or volumetric tetrahedral meshes.

“Mesh Draping” is the clearest fixed-connectivity deformation model. A source triangle or quad mesh $M$ is deformed vertexwise by an MLP with progressive positional encoding, producing $\hat M$ with the same connectivity as $M$; the representation is explicit, artist-centered, and parameterization-free with respect to the target geometry (Hertz et al., 2021). By contrast, NeMo represents each object category by a geometric scaffold and a neural feature field,

$\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$

where $\Gamma_y$ is the mesh and $\Theta_y$ are learned feature vectors attached to vertices (Wang et al., 2021). iNeMo preserves this mesh-plus-feature-field design but extends it to class-incremental learning by adding meshes over time, while DINeMo removes 3D supervision and learns the same style of model from pseudo-correspondences derived from foundation-model features (Fischer et al., 2024, Guo et al., 26 Mar 2025).

A third paradigm treats meshes as sequences. MeshXL introduces Neural Coordinate Field, described as “an explicit coordinate representation with implicit neural embeddings,” and then trains decoder-only autoregressive models over discretized coordinate tokens and special mesh tokens (Chen et al., 2024). MeshMamba replaces transformer-style quadratic attention with Mamba state space models, but still treats each vertex as a token and relies on carefully designed serializations based on template coordinates or DensePose IUV body-part structure (Yoshiyasu et al., 21 Jul 2025).

A fourth family consists of mesh-native graph and equivariant operators. InfoGNN treats faces as graph nodes, constructs KNN graphs in the space of face centers, and combines static geometric features, dihedral-aware local message passing, and dynamic global features in an end-to-end GNN for segmentation and classification (Gao et al., 4 Mar 2025). E(3)-Equivariant Mesh Neural Networks minimally extend EGNN updates with face information via cross products, triangle areas, and normals, then add PointNet++-style hierarchy to handle long-range interactions (Trang et al., 2024). A related line appears in mesh convolutional neural networks for wall shear stress estimation, where graph/mesh convolutions operate directly on CFD surface meshes and predict WSS vectors at mesh vertices (Suk et al., 2021).

A fifth paradigm directly parameterizes connectivity. SpaceMesh defines a continuous latent connectivity space at each vertex and uses it to imply a halfedge mesh representation with guaranteed edge-manifoldness and support for general polygonal meshes (Shen et al., 2024). Instead of predicting faces as discrete index tuples, it predicts adjacency embeddings and permutation embeddings whose thresholding, Sinkhorn normalization, and matching induce a valid $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$0 permutation over halfedges.

Two additional paradigms broaden the mesh-native landscape. DiForm models dynamic meshes implicitly as SDF zero level sets conditioned on an identity code and a deformation code, learning identity-disentangled dynamic surfaces from unregistered scans without template correspondences (Xu et al., 2021). NVMG uses diffusion-generated voxels, converts them to tetrahedral templates, learns a voxel-conditional closest-point network, and then regularizes a deformation process to generate artifact-free volumetric and surface meshes (Zheng et al., 2022).

Paradigm	Core representation	Representative papers
Fixed-connectivity deformation	Vertexwise neural map with preserved source connectivity	(Hertz et al., 2021)
Mesh feature field	Mesh or cuboid scaffold plus learned per-vertex neural features	(Wang et al., 2021, Fischer et al., 2024, Guo et al., 26 Mar 2025)
Coordinate-sequence model	Discretized coordinate tokens or serialized vertex tokens	(Chen et al., 2024, Yoshiyasu et al., 21 Jul 2025)
Mesh-native GNN / equivariant network	Vertex- or face-graph message passing with mesh geometry	(Gao et al., 4 Mar 2025, Trang et al., 2024, Suk et al., 2021)
Continuous connectivity model	Latent adjacency and cyclic halfedge orderings	(Shen et al., 2024)
Implicit dynamic or volumetric mesh model	Identity/deformation-conditioned SDF or tetrahedral deformation pipeline	(Xu et al., 2021, Zheng et al., 2022)

Taken together, these representations show that the defining property of neural mesh models is not a specific network class but a mesh-centric inductive bias: connectivity, polygonal structure, halfedge order, mesh-attached features, or simulation-ready volumetric elements are preserved as first-class design objects.

3. Learning mechanisms and optimization regimes

The optimization regimes vary sharply across paradigms. Mesh Draping is a test-time neural optimizer rather than a dataset-level generator. For each source–target pair it trains a lightweight MLP from scratch and minimizes

$$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$1

where $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$2 combines symmetric Chamfer distance and sparse 3D correspondences, while $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$3 preserves source-mesh angles and local area distributions through an angle term and a Kullback–Leibler divergence term over one-ring area distributions (Hertz et al., 2021). Progressive positional encoding supplies a coarse-to-fine spectral curriculum, which the paper identifies as crucial for stable optimization and detail recovery.

NeMo, iNeMo, and DINeMo instead instantiate analysis-by-synthesis over learned feature fields. NeMo renders vertex-attached neural features through a differentiable renderer and estimates pose by minimizing a probabilistic feature-reconstruction objective rather than regressing pose directly (Wang et al., 2021). Its training couples maximum-likelihood reconstruction with contrastive terms that maximize distances between features at different object parts and between foreground and background, thereby making the loss landscape smoother and more discriminative under occlusion. iNeMo keeps the same render-and-compare core but adds ETF-based latent space initialization, a positional regularization term that keeps classes in predefined latent regions, a continual contrastive loss against unused regions, knowledge distillation over previous meshes, and pose-aware replay for class-incremental learning (Fischer et al., 2024). DINeMo removes 3D annotations entirely and uses bidirectional pseudo-correspondence generation from DINOv2 and Stable Diffusion features: local appearance cues first vote for a global orientation, and that global orientation then constrains local correspondences through visibility (Guo et al., 26 Mar 2025).

Sequence and state-space models use token prediction rather than differentiable rendering. MeshXL factorizes the likelihood of a mesh sequence as standard autoregressive next-token prediction and trains decoder-only transformers with next-token cross-entropy over discretized coordinate tokens (Chen et al., 2024). MeshMamba keeps the tokenized vertex view but replaces transformer blocks with Mamba state space blocks; in MambaDiff3D it predicts diffusion noise on vertices, and in Mamba-HMR it regresses dense whole-body vertices from image-conditioned token features while regularizing edge lengths, Laplacians, and normals (Yoshiyasu et al., 21 Jul 2025).

Direct connectivity models require yet another loss structure. SpaceMesh defines edges by thresholding a spacetime distance over adjacency embeddings,

$$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$4

and supervises them with a cross-entropy over positive and negative vertex pairs. Local cyclic neighbor orderings are represented by permutation embeddings, transformed into soft doubly stochastic matrices via Sinkhorn normalization, and trained with a face loss over ground-truth $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$5-triples in the halfedge structure (Shen et al., 2024).

Mesh-native GNNs operate by message passing rather than rendering or token prediction. InfoGNN uses faces as nodes, constructs edge messages with geometric weighting from face-center distances and dihedral angles, combines static geometry, dynamic local features, and dynamic global features, and trains end-to-end for segmentation or classification (Gao et al., 4 Mar 2025). EMNN extends EGNN by adding face-based invariant messages and normal-based vector updates; because the update equations depend on distances and cross products, E(3)-equivariance follows without spherical harmonics, gauge transport, or spectral operators (Trang et al., 2024).

4. Applications and functional scope

In geometry processing and content creation, neural mesh models support retopology, mesh transfer, and mesh repair. Mesh Draping reuses artist-designed triangle or quad layouts on raw target geometries, including point clouds, polygon soups, and non-manifold meshes, while preserving edge flow, anisotropy, and local density patterns (Hertz et al., 2021). SpaceMesh directly learns manifold polygonal connectivity, generates meshes with tessellation structure learned from dataset populations, and can perform mesh repair by inpainting missing or corrupted regions under point-cloud conditioning (Shen et al., 2024). NVMG extends the scope from surface meshes to volumetric tetrahedral meshes suitable for finite-element-style downstream use (Zheng et al., 2022).

In 3D vision, category-level pose estimation is a central application. NeMo uses mesh-based feature rendering to achieve robust 3D pose estimation under partial occlusion and unseen viewpoints, and it retains strong performance even when the category geometry is approximated by a cuboid rather than a detailed CAD mesh (Wang et al., 2021). iNeMo transfers this mesh-feature-field paradigm to class-incremental learning and also presents the first incremental learning approach for pose estimation (Fischer et al., 2024). DINeMo further shows that neural mesh models can be learned with no 3D annotations by relying on pseudo-correspondences from visual foundation models, while still supporting category-level pose estimation and semantic correspondence (Guo et al., 26 Mar 2025).

In generative modeling and 3D foundation-model design, MeshXL treats meshes as coordinate token sequences and supports unconditional generation, text-to-mesh, image-to-mesh, and partial mesh completion (Chen et al., 2024). MeshMamba adapts state space models to dense articulated meshes, enabling both diffusion-based articulated mesh generation and whole-body human mesh recovery with body, face, and hands at SMPL-X scale (Yoshiyasu et al., 21 Jul 2025). DiForm addresses a different but related problem: reconstructing dynamic hand or body meshes from partial 4D scans while disentangling identity-induced shape variation from pose-induced deformation (Xu et al., 2021).

In discriminative geometric learning, mesh-native GNNs and equivariant mesh networks target classification, segmentation, and physical surrogate modeling. InfoGNN performs mesh classification and segmentation directly on irregular triangle meshes without manifold or watertight restrictions (Gao et al., 4 Mar 2025). EMNN applies E(3)-equivariant message passing with face information to mesh classification and per-vertex segmentation (Trang et al., 2024). Mesh/graph convolutional neural networks for wall shear stress estimation operate directly on the finite-element surface mesh used in CFD and estimate 3D WSS vectors on artery walls, positioning neural mesh models as fast CFD surrogates (Suk et al., 2021).

Outside graphics and vision, mesh-style neural models remain relevant in neuroscience. The two fMRI mesh-learning papers treat spatial or functional voxel neighborhoods as local meshes whose learned edge weights encode information-bearing local dependencies and improve brain-state decoding relative to raw voxel-intensity baselines (Ozay et al., 2012, Onal et al., 2016). This broader usage is methodologically distinct, but it reinforces the idea that mesh structure can function as a learned relational prior rather than merely as a rendering primitive.

5. Empirical findings and comparative behavior

Structure-preserving mesh transfer is one area where explicit neural mesh optimization has shown clear gains. On SHREC-BIM, Mesh Draping reports much lower Dirichlet distortion than HOT and RHM, specifically $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$6 versus $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$7–$$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$8, and a joint quality $$f(\cdot \mid \theta) : \mathbb{R}^3 \to \mathbb{R}^3$$9 of $\Gamma_y$0 versus $\Gamma_y$1–$\Gamma_y$2; on custom mapping tasks it reports $\Gamma_y$3 versus $\Gamma_y$4 for ELC and $\Gamma_y$5 for ISM; against deformation baselines it reports $\Gamma_y$6 versus $\Gamma_y$7 for Neural Cages and $\Gamma_y$8 for ShapeFlow (Hertz et al., 2021).

For category-level pose estimation, NeMo’s robustness under occlusion is the central empirical result. On occluded-PASCAL3D+ at the heaviest occlusion level $\Gamma_y$9, NeMo-MultiCuboid reports $\Theta_y$0, $\Theta_y$1, and median rotation error $\Theta_y$2, compared with StarMap’s $\Theta_y$3, $\Theta_y$4, and $\Theta_y$5, and Res50-Specific’s $\Theta_y$6, $\Theta_y$7, and $\Theta_y$8 (Wang et al., 2021). iNeMo preserves the same category-level mesh prior in the continual setting and reports improvements of $\Theta_y$9–$M$0 in-domain and $M$1–$M$2 in OOD classification; for pose estimation on PASCAL3D it reports $M$3 at $M$4 and $M$5 at $M$6, compared with $M$7 and $M$8 for incremental iCaRL (Fischer et al., 2024). DINeMo, trained with no 3D annotations, reports $M$9 and $\hat M$0 on PASCAL3D+ cars, states that it narrows the gap with fully supervised methods by $\hat M$1, and improves as unlabeled training images increase from $\hat M$2 to $\hat M$3 (Guo et al., 26 Mar 2025).

In direct generative modeling, MeshXL shows that coordinate-token autoregression can compete strongly with mesh-specific transformers and VQ pipelines. On ShapeNet chairs, MeshXL(350M) reports COV $\hat M$4 versus $\hat M$5 for MeshGPT, MMD $\hat M$6 versus $\hat M$7, $\hat M$8-NNA $\hat M$9 versus $M$0, FID $M$1 versus $M$2, and KID $M$3 versus $M$4 (Chen et al., 2024). On Objaverse, scaling from smaller to larger MeshXL models improves COV from $M$5 to $M$6, MMD from $M$7 to $M$8, $M$9-NNA from $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$0 to $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$1, JSD from $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$2 to $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$3, FID from $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$4 to $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$5, and KID from $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$6 to $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$7 (Chen et al., 2024). MeshMamba complements this result by emphasizing dense articulated meshes: on an A100 it reports about $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$8 seconds to generate a $\mathfrak{N}_y = \{\Gamma_y,\Theta_y\},$9-vertex SMPL-X mesh with $\Gamma_y$0 DDIM steps versus $\Gamma_y$1 seconds for a transformer, and Mamba-HMR reports $\Gamma_y$2 FPS with PA-MVE $\Gamma_y$3 and MVE $\Gamma_y$4 on UBody after fine-tuning (Yoshiyasu et al., 21 Jul 2025).

Mesh-native discriminative models also show strong numerical performance. InfoGNN reports $\Gamma_y$5 face accuracy on human body segmentation, $\Gamma_y$6 mIoU on ShapeNetCore part segmentation, and $\Gamma_y$7 accuracy on ModelNet40 classification (Gao et al., 4 Mar 2025). EMNN reports $\Gamma_y$8 on FAUST segmentation and $\Gamma_y$9 on TOSCA classification in its best settings, while remaining significantly faster per epoch than GEM-CNN and EMAN (Trang et al., 2024). In CFD surrogate modeling, mesh convolutional neural networks for wall shear stress estimation process new meshes in less than $\Theta_y$0 seconds, consistently achieve a normalised mean absolute error of $\Theta_y$1, and peak at $\Theta_y$2 median approximation accuracy over the held-out test set (Suk et al., 2021).

6. Limitations, misconceptions, and open research problems

A common misconception is that neural mesh models are synonymous with mesh extraction from implicit fields. The cited literature shows a much broader space. Some models preserve a fixed mesh connectivity throughout optimization; others attach features to a fixed scaffold; others learn connectivity directly; others never output a surface mesh at all but use mesh-like local graphs over fMRI or recurrent states (Hertz et al., 2021, Wang et al., 2021, Shen et al., 2024, Ozay et al., 2012). This suggests that the defining axis is mesh-centric inductive bias, not a single output format.

The trade-off between structure preservation and topological flexibility remains unresolved. Mesh Draping cannot change connectivity and therefore cannot create new vertices or edges for spikes, thin structures, or large semantic mismatches; its authors explicitly note failure cases when source and target differ too strongly in part structure (Hertz et al., 2021). MeshMamba and many dense non-parametric HMR models similarly assume fixed topology and, in the cited experiments, tight clothing rather than large topological departures (Yoshiyasu et al., 21 Jul 2025). By contrast, SpaceMesh and NVMG do generate connectivity, but each retains its own bottlenecks: SpaceMesh currently targets closed manifolds and can still produce self-intersections, while NVMG is limited by $\Theta_y$3 voxel resolution and a non-end-to-end pipeline (Shen et al., 2024, Zheng et al., 2022).

Scalability is another persistent constraint, though it appears in different forms. MeshXL highlights long autoregressive sequences and the lack of explicit topology control as central limitations, motivating future migration toward faster causal sequence architectures such as Mamba, RWKV, or xLSTM (Chen et al., 2024). iNeMo’s continual setting requires CAD models, pose annotations, and a predefined upper bound on the number of classes for ETF allocation (Fischer et al., 2024). DINeMo removes 3D annotations but still assumes a category-specific template and inherits the reliability limits of pseudo-correspondences from foundation-model features (Guo et al., 26 Mar 2025).

For discriminative mesh GNNs, the open problems are less about topology creation and more about geometric fidelity and broader robustness. InfoGNN notes sensitivity to incorrectly modeled data such as flipped face orientation and lacks explicit constraints on segmentation contiguity, which can produce mislabeled islands (Gao et al., 4 Mar 2025). The CFD WSS models show strong surrogate performance, but their cited experiments are performed on synthetic coronary artery datasets and steady-flow settings, which indicates a remaining gap to more general pulsatile and patient-specific regimes (Suk et al., 2021). EMNN demonstrates that simple E(3)-equivariant face-aware message passing can outperform more complicated equivariant mesh architectures on the reported benchmarks, but this does not imply that intrinsic spectral or gauge-based methods are obsolete; rather, it indicates that complexity in mesh neural architectures does not automatically translate into practical performance (Trang et al., 2024).

The overall research direction is therefore not convergence onto one architecture, but refinement of a design space. Current work points toward mesh-native models that can combine direct connectivity learning, efficient long-context sequence modeling, equivariant local operators, neural feature fields for analysis-by-synthesis, and geometry-processing regularizers strong enough to guarantee not only manifoldness but also usable element quality.