MeshMamba: Articulated 3D Mesh Generation

• MeshMamba is a novel family of Mamba-based architectures for articulated 3D mesh learning, employing non-parametric vertex tokenization with structure-aware serialization.
• It transforms dense meshes into 1D vertex sequences using template-based and DensePose IUV ordering, enabling efficient modeling of detailed human body parts like hands, face, and clothing.
• The architecture is instantiated as MambaDiff3D for diffusion-based mesh generation and Mamba-HMR for single-image reconstruction, achieving near-linear complexity and faster inference.

Searching arXiv for the cited MeshMamba paper and closely related Mamba-based 3D/mesh works. MeshMamba is a family of Mamba-based architectures for articulated $3$D mesh learning, introduced for two tasks: generation of plausible articulated meshes and reconstruction of a human mesh from a single RGB image. Its defining idea is to treat a mesh as a sequence of vertex tokens while imposing a structure-aware serialization derived either from DensePose body-part IUV annotations or from template-mesh vertex coordinates. This design targets dense whole-body meshes at full SMPL and SMPL-X resolution, including clothing, hands, and face, and is instantiated as MambaDiff3D for diffusion-based mesh generation and Mamba-HMR for single-image human mesh recovery (Yoshiyasu et al., 21 Jul 2025).

1. Problem setting and conceptual scope

MeshMamba addresses articulated $3$D meshes in a non-parametric vertex representation. In the formulation used by the method, a mesh has $N$ vertices, $F$ triangular faces, and vertex coordinates

$\mathbf{x} \in \mathbb{R}^{N \times 3}.$

A template mesh $\mathcal{M}_0$ is given in a canonical pose, and all training meshes share the same topology and point-to-point correspondence with that template. This correspondence assumption is central because it makes per-vertex tokenization consistent across the dataset (Yoshiyasu et al., 21 Jul 2025).

The motivation is rooted in the limitations of earlier dense human mesh modeling. Parametric methods regress SMPL or SMPL-X coefficients, whereas non-parametric methods directly predict vertex positions. The paper emphasizes that vertex-based representations are more directly compatible with neural networks, more general for geometric deformation, and better suited to clothing and non-body-specific detail. However, when every vertex becomes a token, dense articulated meshes reach thousands or over ten thousand tokens, and transformer backbones become difficult to use because standard self-attention has quadratic complexity in sequence length. Previous vertex-transformer HMR systems were therefore typically limited to around $500$ vertices and relied on an upsampling stage, which tends to blur or miss local geometry (Yoshiyasu et al., 21 Jul 2025).

Whole-body modeling intensifies this difficulty. The paper demonstrates full-resolution SMPL meshes with $6{,}890$ vertices and full-resolution SMPL-X meshes with $10{,}475$ vertices. At this scale, coarse body-only meshes lose finger articulation, omit face detail, and inadequately capture clothing wrinkles or garment-specific shape. MeshMamba is positioned as a response to that bottleneck: it seeks to make dense articulated whole-body modeling practical by replacing transformer-based vertex processing with Mamba State Space Models, which the paper characterizes as near-linear in sequence length and efficient at training and inference (Yoshiyasu et al., 21 Jul 2025).

Component	Task	Key characteristics
MeshMamba	Backbone	Mamba-based sequence modeling over serialized vertex tokens
MambaDiff3D	Generation	Denoising diffusion for articulated $3$D mesh generation
Mamba-HMR	Reconstruction	Single-image whole-body mesh recovery

A common misconception is to treat MeshMamba as simply “Mamba on meshes.” The paper’s own framing is narrower and more specific: the decisive ingredient is not merely replacing attention with an SSM, but making that substitution work for articulated surfaces through a serialization that respects articulated structure.

2. Vertex serialization and articulated structure

MeshMamba is order-sensitive because Mamba is not permutation-invariant over tokens. The method therefore converts mesh vertices into a $3$0D sequence using a template-derived permutation that is reused for all examples. The paper presents two main serialization strategies (Yoshiyasu et al., 21 Jul 2025).

The first is template-vertex-location-based ordering. Template vertices in a canonical pose, typically a T-pose, are lexicographically sorted by their $3$1D coordinates. The six explicit orderings are xyz, -xyz, yzx, -yzx, zxy, and -zxy, obtained by varying axis order and sign. This creates several plausible traversals of the articulated surface while preserving a consistent template-based scan.

The second is DensePose body-part IUV-based ordering. DensePose provides a body-part index $3$2 and UV coordinates $3$3. Serialization first sorts parts by centroid coordinates using the same six axis/sign variants, then sorts vertices within each part by $3$4 and $3$5. The stated effect is semantic clustering of torso, arm, hand, and other anatomically related vertices into contiguous sequence regions.

The rationale is that Mamba performs a selective sequential scan. If physically or semantically related vertices are adjacent in the serialized sequence, the recurrent state updates can more naturally model limb continuity, body-part-specific deformations, hand geometry, and clothing attached to specific regions. The ablations reported in the paper support this interpretation: random ordering caused MeshMamba to fail to learn properly; the default SMPL connectivity order performed noticeably worse; structure-aware orderings improved results; and combining two orderings yielded the best quality-efficiency trade-off (Yoshiyasu et al., 21 Jul 2025).

The paper also adopts a practical compromise on multi-ordering. Although multiple serializations across many layers can expose different adjacency patterns, changing ordering at every layer requires gather and index operations that slow the network. The reported design therefore uses one serialization strategy for most Mamba layers and a second strategy in only one layer. In the serialization ablation, SMPL ×1 + XYZ ×1 achieved the best reported $3$6-NNA score of $3$7, compared with $3$8 for SMPL connectivity ×1, $3$9 for part-IUV ×1, and $N$0 for SMPL ×1 + IUV ×1 (Yoshiyasu et al., 21 Jul 2025).

The method is explicitly contrasted with point-cloud Mamba approaches that use space-filling curves such as z-order or Hilbert curves. The paper regards those as unsuitable here because articulated meshes deform heavily, and because MeshMamba relies on fixed topology and template correspondence rather than free-form point sets.

3. State-space formulation and backbone architecture

MeshMamba is built on the standard state-space formulation

$N$1

which is then turned into a selective, input-dependent system in Mamba: $N$2 The paper emphasizes that this selective mechanism allows context-aware remembering and forgetting while retaining near-linear efficiency through a hardware-aware parallel scan (Yoshiyasu et al., 21 Jul 2025).

At the architectural level, MeshMamba uses a standard Mamba block comprising a selective SSM layer, linear layers, a convolution layer, and nonlinear activation layers. The architecture figure includes explicit feature permutation based on serialized tokens, so sequence reindexing is an explicit part of the pipeline. Each mesh vertex is a token. For generation, tokens begin as noisy $N$3D coordinates $N$4. For reconstruction, each token is a query feature associated with a mesh vertex and fused with image features and position embedding (Yoshiyasu et al., 21 Jul 2025).

Both MambaDiff3D and Mamba-HMR employ a U-shaped or encoder-decoder-like stacked arrangement with a shallow group of $N$5 Mamba blocks, a mid block, and a deep group of $N$6 Mamba blocks, connected by skip connections from shallow to deep blocks. The paper states that total depth is $N$7 blocks if the middle block is counted, but it does not specify the exact value of $N$8 or the feature dimension $N$9 numerically in the provided text. Likewise, it does not specify normalization type beyond what is implicit in the standard Mamba block, does not provide exact activation names, and does not state that MeshMamba uses bidirectional Mamba (Yoshiyasu et al., 21 Jul 2025).

The paper also notes what MeshMamba is not. It does not introduce a new symbolic equation named “MeshMamba,” does not define a new pose-angle latent parameterization, and does not rely on a low-dimensional explicit parametric body latent. Instead, it uses correspondence-based template deformation and dense per-vertex sequence processing.

4. MambaDiff3D and dense articulated mesh generation

MambaDiff3D is the generative instantiation of MeshMamba. It is a denoising diffusion model inspired by U-ViT-style backbones, but with transformer blocks replaced by Mamba blocks. Its input is noisy mesh coordinates

$F$0

and its output is predicted noise

$F$1

An input MLP maps $F$2 to token embeddings, the sequence is processed by the shallow-mid-deep Mamba stack with skip connections, and an output MLP predicts the denoising target. A time embedding for the diffusion timestep $F$3 is added to every Mamba block by summation (Yoshiyasu et al., 21 Jul 2025).

The model uses v-prediction parameterization with a cosine variance scheduler. The training objective is

$F$4

Sampling uses DDIM with $F$5 diffusion timesteps and reported sampling steps of $F$6, $F$7, and $F$8. The paper reports that reasonable-quality $F$9-vertex meshes can be generated in about $\mathbf{x} \in \mathbb{R}^{N \times 3}.$0 second with $\mathbf{x} \in \mathbb{R}^{N \times 3}.$1 DDIM steps, and that generation of a $\mathbf{x} \in \mathbb{R}^{N \times 3}.$2-vertex mesh with $\mathbf{x} \in \mathbb{R}^{N \times 3}.$3 DDIM steps on an NVIDIA A100 takes about $\mathbf{x} \in \mathbb{R}^{N \times 3}.$4 seconds, compared with $\mathbf{x} \in \mathbb{R}^{N \times 3}.$5 seconds for a transformer backbone using PyTorch Flash Attention. On a V100, the reported comparison is $\mathbf{x} \in \mathbb{R}^{N \times 3}.$6 seconds versus $\mathbf{x} \in \mathbb{R}^{N \times 3}.$7 seconds. For training with $\mathbf{x} \in \mathbb{R}^{N \times 3}.$8 vertex tokens, the reported epoch time is about $\mathbf{x} \in \mathbb{R}^{N \times 3}.$9 minutes for MeshMamba versus $\mathcal{M}_0$0 minutes for the transformer, under the same $\mathcal{M}_0$1 NVIDIA V100 setting with batch size $\mathcal{M}_0$2 (Yoshiyasu et al., 21 Jul 2025).

A notable design detail concerns local geometric quality. The paper argues that vertex-only generation tends to produce local noise, whereas Jacobian generation can produce global distortion from accumulated tangential errors. MambaDiff3D therefore generates both vertex positions and surface normals and combines them through a Poisson reconstruction-style postprocess. For triangle $\mathcal{M}_0$3, the gradient is computed as

$\mathcal{M}_0$4

where $\mathcal{M}_0$5 is the Jacobian of the generated vertices after smoothing and $\mathcal{M}_0$6 is the relative rotation between generated normals and normals derived from smoothed vertices. The resulting gradients are inserted into a Poisson system to reconstruct a globally consistent mesh. The paper explicitly states that this Poisson solve is not end-to-end and that no gradients flow back from the solver.

Setting	1-NNA $\mathcal{M}_0$7	Other reported metrics
MambaDiff3D, SURREAL train set	53.1	FID $\mathcal{M}_0$8, APD $\mathcal{M}_0$9
MambaDiff3D, AMASS train set	55.1	FID $500$0, APD $500$1

These results are reported on the SURREAL test set of $500$2 meshes, using $500$3-NNA, FID computed from joint locations, and APD also computed from joint locations. The paper emphasizes the best $500$4-NNA among the compared methods and a strong diversity-realism trade-off, while also noting that VPoser has lower FID but substantially lower APD and worse $500$5-NNA. Qualitatively, the method is reported to generate dense human meshes in clothes, with grasping hands, and to generalize to other mammals using Animal3D (Yoshiyasu et al., 21 Jul 2025).

5. Mamba-HMR and single-image whole-body reconstruction

Mamba-HMR adapts MeshMamba to human mesh recovery from a single image. Its stated contribution is to extend non-parametric vertex regression from earlier body-only, low-resolution settings to dense whole-body meshes including body, hands, and face. It follows the general design of Mesh Transformer but replaces transformer blocks with MeshMamba blocks (Yoshiyasu et al., 21 Jul 2025).

The image encoder is HRNet-W48 with input image size $500$6 and an extracted feature map of $500$7. The backbone is initialized from weights pretrained for $500$8D human pose detection on COCO-WholeBody. Image features are fed into the Mamba stack through body joint queries and mesh vertex queries together with position embedding. The paper does not provide an explicit fusion equation, but it states that the output is a full-resolution mesh produced directly, without the obligatory low-resolution prediction and upsampling used in earlier non-parametric methods. It also states that this substantially reduces model parameters (Yoshiyasu et al., 21 Jul 2025).

The total training loss is

$500$9

In addition to vertex and joint supervision, the method introduces three local geometric regularizers: Laplacian loss based on mean curvature normals derived from the cotangent Laplacian matrix, edge loss on edge lengths, and normal loss on face normals. The paper emphasizes that these local geometric losses are vital for preserving dense local shape quality (Yoshiyasu et al., 21 Jul 2025).

Method	UBody overall accuracy	FPS
Mamba-HMR	PA-MVE All $6{,}890$0, MVE All $6{,}890$1	22
Mamba-HMR fine-tuned	PA-MVE All $6{,}890$2, MVE All $6{,}890$3	22

On UBody, the paper reports that Mamba-HMR achieves the best overall PA-MVE and MVE among the listed methods. The fine-tuned model reports PA-MVE All $6{,}890$4, Hands $6{,}890$5, Face $6{,}890$6, and MVE All $6{,}890$7, Hands $6{,}890$8, Face $6{,}890$9, at $10{,}475$0 FPS. The interpretation in the paper is explicit: the method is especially strong on overall whole-body accuracy and face MVE, but it is not the best on every hand metric; AiOS has the best PA-MVE for hands, and Multi-HMR-B has the best hand MVE. The reported inference speed of about $10{,}475$1 FPS is described as near real-time, faster than OSX-L at $10{,}475$2 FPS, slower than NLF-L at $10{,}475$3 FPS, and roughly similar to SMPLer-X-L at $10{,}475$4 FPS and Multi-HMR-B at $10{,}475$5 FPS (Yoshiyasu et al., 21 Jul 2025).

6. Empirical standing, limitations, and relation to adjacent work

The empirical case for MeshMamba rests on both quality and scalability. In the network-block ablation for generation, MLP and GNN backbones perform much worse than sequence models, with reported $10{,}475$6-NNA values of $10{,}475$7 for MLP and $10{,}475$8 for GNN, compared with $10{,}475$9 for a transformer and $3$0 for Mamba. This supports the paper’s claim that transformer and Mamba are both strong, but that Mamba is slightly better in quality while being much faster in dense mesh diffusion (Yoshiyasu et al., 21 Jul 2025).

The method also has clear limits. The paper explicitly states a fixed-topology and tight-clothing limitation: MeshMamba is limited to tight clothing with fixed topology. It also identifies a generalization gap to unseen datasets relative to large-scale whole-body systems trained on broader data, such as SMPLer-X. Qualitative failure modes include local noise when geometric regularization is weak, distortions under poorer serialization choices, and missing hand detail when downsampled meshes are used. A plausible implication is that the backbone’s efficiency does not remove the classical dependence of correspondence-based mesh learning on strong template assumptions.

The name also requires disambiguation. “MeshMamba: State Space Models for Articulated 3D Mesh Generation and Reconstruction” (Yoshiyasu et al., 21 Jul 2025) is distinct from “Mesh Mamba: A Unified State Space Model for Saliency Prediction in Non-Textured and Textured Meshes” (Zhang et al., 2 Apr 2025), which predicts per-face saliency from geometry and optionally UV-aligned texture, using subgraph embedding and a bidirectional SSM. The two works share the use of Mamba-style state-space modeling on mesh data, but they target different problems and use different tokenizations: vertex sequences for articulated mesh modeling in the former, topology-preserving face subgraphs for saliency prediction in the latter.

Within the broader Mamba-for-$3$1D landscape, MeshMamba is also adjacent to methods that are not themselves articulated mesh generators. “3DMambaComplete” (Li et al., 2024) is a point cloud completion network whose decoder deforms local $3$2D grid or mesh patches around HyperPoints into dense $3$3D point sets, but its output remains a point cloud rather than a triangular mesh. “SpineMamba” (Zhang et al., 2024) is a voxel-based spinal segmentation model inside the nnU-Net framework, where Mamba layers capture long-range spatial dependencies and a Vertebrae Shape Priors module injects anatomical structure. These comparisons suggest a broader pattern across recent $3$4D Mamba work: local geometry-biased operators are preserved, while long-range propagation is delegated to linear-time state-space sequence modeling.

In that context, MeshMamba’s specific contribution is to show that dense articulated whole-body geometry can be modeled directly at SMPL and SMPL-X vertex resolution when the sequence order is made anatomically and structurally meaningful. Its significance follows from that conjunction of design choices rather than from Mamba alone: sequence modeling, template correspondence, and structure-aware serialization are all necessary parts of the method’s reported performance and efficiency.