MambaDiff3D: Diffusion for 3D Mesh Generation

• MambaDiff3D is a diffusion-based model within the MeshMamba framework that generates dense, articulated 3D meshes using Mamba State Space Models and structured vertex serialization.
• It replaces transformer blocks with efficient Mamba-SSMs and serializes mesh vertices into 1D token sequences to overcome scalability issues and preserve fine geometric details.
• Benchmark results show state-of-the-art performance with metrics such as a 1-NNA of 53.1, FID of 0.32, and APD of 23.01 on SURREAL, while significantly reducing generation time.

Searching arXiv for the specified paper and closely related Mamba-based 3D generation work. arXiv API query: search "MeshMamba articulated 3D mesh generation MambaDiff3D" MambaDiff3D is a denoising diffusion model for dense 3D articulated mesh generation introduced as part of the MeshMamba framework. It is designed to generate full-resolution articulated meshes by replacing transformer blocks with Mamba State Space Models (Mamba-SSMs) and by serializing mesh vertices into structure-aware 1D token sequences. Within MeshMamba, MambaDiff3D serves the generative branch, while Mamba-HMR addresses 3D human mesh recovery from a single image. The model is motivated by the poor scalability of transformer-based vertex models as mesh resolution increases and is reported to generate dense 3D human meshes, including clothed bodies and grasping hands, while being substantially faster than transformer diffusion baselines (Yoshiyasu et al., 21 Jul 2025).

1. Definition and placement within MeshMamba

MeshMamba is a neural network framework for learning 3D articulated mesh models with Mamba-SSMs, and MambaDiff3D is its diffusion-based mesh generation component (Yoshiyasu et al., 21 Jul 2025). The framework has two main applications: MambaDiff3D for unconditional or class-conditional 3D articulated mesh generation, and Mamba-HMR for 3D human mesh recovery from a single image (Yoshiyasu et al., 21 Jul 2025).

The central problem addressed by MambaDiff3D is the generation of dense articulated meshes with thousands of vertices. Prior transformer-based vertex models are described as scaling poorly because self-attention is quadratic in the number of vertex tokens, which makes full-resolution mesh generation impractical. This limitation is especially consequential for hands, face, and clothing geometry, since coarse meshes cannot accurately represent fingers, grasp configurations, or fine surface detail. MambaDiff3D is introduced to overcome this bottleneck by using Mamba-SSMs, which preserve near-linear complexity while remaining effective on long token sequences (Yoshiyasu et al., 21 Jul 2025).

The model is reported to handle meshes at SMPL resolution with 6890 vertices and SMPL-X resolution with 10,475 vertices (Yoshiyasu et al., 21 Jul 2025). It is shown generating full-body human meshes, clothed bodies, grasping hand poses, and meshes for other mammals when trained on Animal3D (Yoshiyasu et al., 21 Jul 2025). This suggests that its main technical significance lies in making dense articulated mesh diffusion practical at token counts that were previously dominated by scalability constraints.

2. State-space formulation and architectural design

MambaDiff3D is built on the standard state-space model formulation reviewed in MeshMamba. In continuous time, the model uses

$$h'(t) = {\bf A} h(t) + {\bf B} x(t), \qquad y(t) = {\bf C} h(t)$$

and after discretization with zero-order hold,

$$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$

with

$$\bar{\bf A} = {\rm exp}(\Delta {\bf A}), \qquad \bar{\bf B} = (\Delta {\bf A})^{-1}\left({\rm exp}(\Delta {\bf A}) - I\right)\cdot \Delta {\bf B}, \qquad \bar{\bf C} = {\bf C}.$$

The same formulation is also written in convolutional form as

$$\bar{\bf K} = (\bar{\bf C}\bar{\bf B}, \bar{\bf C}\bar{\bf A}\bar{\bf B}, \ldots, \bar{\bf C}\bar{\bf A}^{k}\bar{\bf B}, \ldots), \qquad y = x * \bar{\bf K}.$$

Mamba generalizes this to a linear time-variant selective system,

$$h'(t) = \bar{\bf A}(x(t))h(t) + \bar{\bf B}(x(t))x(t), \qquad y(t) = {\bf C}(x(t))h(t),$$

so that the model can selectively remember or forget information based on the input (Yoshiyasu et al., 21 Jul 2025).

Within MeshMamba, a standard Mamba block consists of a selective SSM layer, linear layers, a convolution layer, and a nonlinear activation (Yoshiyasu et al., 21 Jul 2025). MambaDiff3D adopts a U-ViT-style diffusion network in which attention blocks are replaced by these Mamba blocks. The input is a noisy vertex sequence

${\bf x}_t \in \mathbb{R}^{N \times 3},$

and the output is predicted noise

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

The network contains an input MLP that maps vertex coordinates into a $d$-dimensional latent embedding, followed by $L+1$ Mamba-block layers arranged as a shallow first half, a mid block, and a deep second half, with skip connections from early blocks to later blocks, and an output MLP that maps features back to 3D noise predictions. The diffusion timestep embedding is added to every Mamba block by summation (Yoshiyasu et al., 21 Jul 2025).

A plausible implication is that MambaDiff3D is best understood not as a geometric decoder operating on explicit mesh neighborhoods, but as a long-sequence denoiser over serialized vertices whose effectiveness depends critically on sequence ordering.

3. Vertex tokenization and serialization

A defining component of MambaDiff3D is vertex serialization. MeshMamba assumes a template mesh $\mathcal{M}_0$ in canonical pose, training meshes $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$0, and shared connectivity with known vertex correspondences. Each mesh is represented as vertex positions

$$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$1

Because Mamba is sequence-order sensitive rather than permutation-invariant, the conversion of a mesh into a 1D token sequence is a core modeling decision (Yoshiyasu et al., 21 Jul 2025).

The paper explores two main serialization strategies. The first is template coordinate sorting, in which vertices are ordered according to the 3D coordinates of a template mesh, using lexicographic orderings such as xyz, -xyz, yzx, -yzx, zxy, and -zxy. The second is DensePose body-part IUV sorting, which first sorts by body-part index $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$2, then by $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$3, then by $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$4, while body-part centroids are themselves ordered with the same six axis variants (Yoshiyasu et al., 21 Jul 2025).

The paper reports that random ordering fails, mesh connectivity alone performs poorly, DensePose-based serialization is better, and combining two serialization strategies works best (Yoshiyasu et al., 21 Jul 2025). For efficiency, one serialization strategy is typically used for all but one Mamba layer, with a second strategy inserted in one layer because continually changing order across layers incurs gather and indexing overhead (Yoshiyasu et al., 21 Jul 2025). This serialization mechanism is therefore not incidental preprocessing; it is a structural requirement for making Mamba effective on articulated meshes.

In broader Mamba-based 3D modeling, sequence ordering repeatedly appears as a decisive factor. Related work on dense volumetric medical segmentation also finds that scan direction and permutation strongly affect Mamba performance, leading to bidirectional or multi-directional schemes rather than naive flattening (Lumetti et al., 2024, Wang et al., 25 Mar 2025). This suggests a broader design principle: Mamba-based 3D models often succeed only when the serialization procedure preserves enough geometric or anatomical structure for selective state updates to remain meaningful.

4. Diffusion objective and mesh reconstruction pipeline

MambaDiff3D uses a standard diffusion setup with v-prediction parameterization and a cosine variance schedule. The training loss is

$$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$5

with weighting

$$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$6

The model is trained to predict Gaussian noise from noisy vertex coordinates (Yoshiyasu et al., 21 Jul 2025).

At inference time, it uses DDIM sampling. The paper sets total diffusion steps to $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$7 and tests 50, 100, and 250 sampling steps (Yoshiyasu et al., 21 Jul 2025). The paper states that even with fewer steps, MambaDiff3D remains fast and yields reasonable meshes (Yoshiyasu et al., 21 Jul 2025).

A notable design choice is that generation is not based solely on raw vertex coordinates. The paper argues that dense meshes generated directly from vertices can exhibit local noise, whereas Jacobian-based representations can distort globally. To preserve fine detail, MambaDiff3D generates and combines vertex positions and surface normals, then reconstructs a smooth mesh using a Poisson system in the gradient domain (Yoshiyasu et al., 21 Jul 2025). The process consists of smoothing the generated vertices, combining the smoothed geometry with predicted normals, computing triangle gradients by

$$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$8

where $$h'(t) = \bar{\bf A} h(t) + \bar{\bf B} x(t), \qquad y(t) = \bar{\bf C} h(t)$$9 is the Jacobian of the smoothed generated vertices and $$\bar{\bf A} = {\rm exp}(\Delta {\bf A}), \qquad \bar{\bf B} = (\Delta {\bf A})^{-1}\left({\rm exp}(\Delta {\bf A}) - I\right)\cdot \Delta {\bf B}, \qquad \bar{\bf C} = {\bf C}.$$0 is the relative rotation aligning generated normals with those from smoothed vertices, and then solving a Poisson system to stitch gradients into a whole mesh (Yoshiyasu et al., 21 Jul 2025). The right-hand side is constant for meshes with fixed connectivity, so the factorization can be reused; however, the Poisson reconstruction is not end-to-end, and gradients do not flow back through the Poisson solve during training (Yoshiyasu et al., 21 Jul 2025).

This suggests that MambaDiff3D couples a token-sequence generative prior with a geometry-aware post-processing stage. Rather than requiring the denoiser itself to guarantee globally smooth surfaces, it delegates part of that burden to a downstream gradient-domain reconstruction procedure.

5. Training data, evaluation protocol, and reported results

For MambaDiff3D, training uses a mixture of articulated mesh datasets: SURREAL, DFAUST, CAPE, GRAB, AMASS, BARC, and Animal3D (Yoshiyasu et al., 21 Jul 2025). Meshes are preprocessed so that global positions and orientations are aligned at the root (Yoshiyasu et al., 21 Jul 2025). Optimization uses Adam with initial learning rate $$\bar{\bf A} = {\rm exp}(\Delta {\bf A}), \qquad \bar{\bf B} = (\Delta {\bf A})^{-1}\left({\rm exp}(\Delta {\bf A}) - I\right)\cdot \Delta {\bf B}, \qquad \bar{\bf C} = {\bf C}.$$1, reduced by a factor of 10 after half the epochs. Training is conducted either on a single cluster node with 8 NVIDIA A100 GPUs or on 6 nodes with 4 NVIDIA V100 GPUs (Yoshiyasu et al., 21 Jul 2025).

Evaluation for 3D human generation is performed on the SURREAL test set with 200 meshes, using 1-NNA, FID, and APD. The main reported generation results are summarized below (Yoshiyasu et al., 21 Jul 2025).

Method	1-NNA	Notes
Pose-NDF	92.0	SURREAL
NRDF	81.6	SURREAL
VPoser	60.7	SURREAL
Parametric Diffusion	59.6	SURREAL
DiffSurf	54.4	SURREAL
MambaDiff3D (SURREAL)	53.1	best reported 1-NNA
MambaDiff3D (AMASS)	55.1	trained on AMASS

MambaDiff3D also reports FID = 0.32 on SURREAL and APD = 23.01 on SURREAL (Yoshiyasu et al., 21 Jul 2025). The paper interprets these numbers as indicating that the model generates shapes that are more realistic and more diverse than prior pose-based and vertex-based baselines (Yoshiyasu et al., 21 Jul 2025).

The qualitative findings reported in the paper include realistic articulated body poses, clothing deformations, hand grasping configurations, and dense full surfaces without the finger loss seen in downsampled-mesh methods (Yoshiyasu et al., 21 Jul 2025). The paper compares MambaDiff3D favorably to DiffSurf, PoseNDF, and NRDF in this regard (Yoshiyasu et al., 21 Jul 2025).

6. Ablations, efficiency, and relation to adjacent Mamba-based 3D diffusion models

The paper includes several ablations that identify the main performance drivers. Replacing Mamba with other block types degrades performance. The reported 1-NNA values are 73.7 for MLP, 74.2 for GNN, 53.6 for Transformer, and 53.1 for Mamba (Yoshiyasu et al., 21 Jul 2025). Serialization is also shown to be crucial: SMPL connectivity ×1 gives 60.0, part-IUV ×1 gives 54.4, part-IUV ×7 gives 53.0, and SMPL ×1 + XYZ ×1 gives 53.1 (Yoshiyasu et al., 21 Jul 2025). The paper concludes that a combination of two serialization strategies offers a good balance of quality and efficiency (Yoshiyasu et al., 21 Jul 2025).

The mesh representation ablations further report that generating vertices alone causes local noise, Jacobian-based representations can distort globally, and combining vertex positions and normals improves smoothness and detail preservation (Yoshiyasu et al., 21 Jul 2025). The paper also explicitly states that generating on downsampled meshes loses fingers and cannot recover hand shapes, which motivates full-resolution generation (Yoshiyasu et al., 21 Jul 2025).

Efficiency is a central part of the model’s contribution. Training a 6890-vertex MeshMamba reportedly takes about 18 minutes per epoch on $$\bar{\bf A} = {\rm exp}(\Delta {\bf A}), \qquad \bar{\bf B} = (\Delta {\bf A})^{-1}\left({\rm exp}(\Delta {\bf A}) - I\right)\cdot \Delta {\bf B}, \qquad \bar{\bf C} = {\bf C}.$$2 V100s with batch size 8, compared with 100 minutes for the transformer baseline (Yoshiyasu et al., 21 Jul 2025). For a 10,475-vertex mesh, on A100 hardware, Mamba takes about 4.5 s with 250 DDIM steps, compared with 28.1 s for the transformer baseline; on V100, the corresponding numbers are 6.6 s and 58.3 s (Yoshiyasu et al., 21 Jul 2025). With only 50 DDIM steps, MambaDiff3D can generate reasonable quality meshes in about 1 second (Yoshiyasu et al., 21 Jul 2025).

These results place MambaDiff3D within a broader class of Mamba-based 3D generative systems that use state-space sequence modeling to replace or reduce attention. For example, DiM-3D uses Mamba with bidirectional SSMs for 3D shape generation on voxelized point clouds and reports improved scalability relative to DiT-3D (Mo, 2024). Skip Mamba Diffusion applies a Mamba-based latent diffusion denoiser to monocular 3D semantic scene completion (Liang et al., 13 Jan 2025). Although these works operate on different 3D representations—voxels, point clouds, latent scenes, or articulated meshes—they share the premise that selective state-space models are better suited than attention-heavy backbones for very long 3D sequences (Mo, 2024, Liang et al., 13 Jan 2025).

A common misconception would be to treat MambaDiff3D as merely a direct replacement of transformer blocks with Mamba layers. The MeshMamba results indicate that this would omit the most consequential design choice: structure-aware vertex serialization. The ablations show that ordering strategy materially affects performance, and random or weakly structured orderings perform substantially worse (Yoshiyasu et al., 21 Jul 2025). A plausible implication is that the scalability advantage of Mamba becomes useful for 3D mesh generation only when the tokenization itself preserves articulated structure.

The paper notes two limitations: the method is still limited to tight clothing with fixed topology, and generalization to unseen datasets is weaker than methods trained on broader multi-dataset data (Yoshiyasu et al., 21 Jul 2025). These limitations indicate that MambaDiff3D advances dense articulated mesh generation primarily within the regime of fixed-topology template-based modeling rather than unrestricted mesh synthesis.