MeshTailor: Seam Generation & Human Mesh Estimation

• MeshTailor is a dual-purpose system that generates edge-aligned UV seams directly on 3D meshes while also enabling consistent human mesh estimation.
• It employs a dual-stream encoder combining topological (GraphSAGE-based) and global geometric (point-cloud) contexts, orchestrated through an autoregressive Transformer.
• In human mesh estimation, MeshTailor maps anthropometric measurements to SMPL-X shape parameters, ensuring stable body shapes across video frames.

Searching arXiv for the cited MeshTailor paper and closely related seam/UV methods to ground the article. In recent arXiv literature, MeshTailor denotes two distinct mesh-centric systems. The 2026 paper "MeshTailor: Cutting Seams via Generative Mesh Traversal" defines a mesh-native generative framework for synthesizing edge-aligned seams on 3D surfaces, with the explicit aim of producing UV seam chains directly on the mesh graph for coherent chart cutting and downstream unwrapping (Ma et al., 28 Mar 2026). Separately, the 2024 paper "Leveraging Anthropometric Measurements to Improve Human Mesh Estimation and Ensure Consistent Body Shapes" uses MeshTailor to name a pipeline for producing consistent human body shapes across video frames by mapping anthropometric measurements to SMPL-X shape parameters and combining them with 3D pose via inverse kinematics (Ludwig et al., 2024). The shared theme is direct manipulation of mesh-relevant structure, but the two systems address different technical problems: seam layout for surface parameterization, and stable body-shape recovery for human mesh estimation.

1. Mesh-native seam generation on 3D surfaces

In the seam-generation sense, MeshTailor is formulated as a system for automatically placing UV seams directly on a 3D mesh so that the surface can be cut into coherent charts and flattened into 2D with minimal fragmentation and artist-friendly structure. The paper frames this as a seam-generation problem rather than a pure parameterization problem: the target output is not merely a UV map with acceptable distortion, but an explicit set of edge-aligned seam chains drawn on the mesh surface itself. These seams are intended to follow natural shape boundaries, remain continuous, and preferably form long loops or clean chains that match how experienced artists would place cuts (Ma et al., 28 Mar 2026).

The input mesh is written as $\mathcal{M}=(\mathcal{V},\mathcal{E},\mathcal{F})$, with vertex set $\mathcal{V}$, edge set $\mathcal{E}$, and face set $\mathcal{F}$. A seam chain is represented as a walk

$\mathbf{c}=(v_1,\dots,v_T)$

subject to the adjacency constraint

$$(v_t,v_{t+1})\in \mathcal{E}, \quad \forall t\in\{1,\dots,T-1\}.$$

Open chains and closed loops are both supported; a loop is the special case $v_1=v_T$. The seam layout may also be viewed as an edge subset $\mathcal{S}\subseteq\mathcal{E}$, or equivalently as binary edge labels $y_e\in\{0,1\}$ with $y_e=\mathbb{I}[e\in\mathcal{S}]$.

This formulation is significant because it makes the output directly usable for cutting UV charts. The seams are not predicted as free-space curves or image-space regions and then projected back; they are generated as vertex walks on the mesh itself. A plausible implication is that the representation makes seam editing and downstream atlas construction more robust, since the prediction is already intrinsic to the target discretization.

2. ChainingSeams, dual-stream encoding, and autoregressive traversal

A central technical contribution is ChainingSeams, a hierarchical serialization of the seam graph designed to make autoregressive generation compatible with structured seam layouts. Because seam graphs are unordered while sequence models require ordered targets, the method decomposes the seam set into seam chains and serializes them using a loops-first, balance-first, large-patch-first rule. The ordering starts from a set of patches $\mathcal{V}$0, repeatedly selects the largest patch,

$\mathcal{V}$1

finds internal loop candidates in that patch, and chooses the loop that most evenly splits it according to

$\mathcal{V}$2

After all loops are ordered, the remaining open chains are appended deterministically, for example by decreasing chain length (Ma et al., 28 Mar 2026).

MeshTailor uses a dual-stream encoder because seam placement depends on both topological context and geometric context. The topological stream processes the mesh as a graph with GraphSAGE. Each vertex first receives coordinate-based features: a Fourier feature encoder $\mathcal{V}$3 is applied to vertex data $\mathcal{V}$4 consisting of position plus normal, and the result is concatenated with the raw coordinates and passed through an MLP,

$\mathcal{V}$5

GraphSAGE layers then perform message passing,

$\mathcal{V}$6

yielding connectivity-aware embeddings that are fused with raw point features:

$\mathcal{V}$7

The geometric stream extracts global shape context from a surface point cloud using a frozen pretrained point-cloud encoder, Michelangelo, which produces a token set

$\mathcal{V}$8

These two streams are fused by cross-attention,

$\mathcal{V}$9

so that the final per-vertex embedding contains both local mesh structure and global shape semantics.

The decoder is a Transformer-based autoregressive model with a mesh-native pointer layer. At decoding step $\mathcal{E}$0, with hidden state $\mathcal{E}$1, the next-vertex distribution is

$\mathcal{E}$2

where $\mathcal{E}$3 is learned and $\mathcal{E}$4 enforces validity. The candidate universe is

$\mathcal{E}$5

The mask is defined so that $\mathcal{E}$6 if $\mathcal{E}$7 is in the 1-ring neighborhood of the current vertex $\mathcal{E}$8, or if it is a valid control token such as [EOC] or [EOS]; otherwise $\mathcal{E}$9. This makes every generated step a legal mesh-edge traversal and ensures edge alignment by construction.

3. Training objective, inference procedure, and empirical evaluation

Training uses a standard autoregressive negative log-likelihood over serialized seam tokens:

$\mathcal{F}$0

where the target sequence $\mathcal{F}$1 consists of vertex identifiers together with the special tokens [EOC] and [EOS]. The decoder uses rotary positional embeddings (RoPE) and a chain-local positional embedding that resets after every [EOC], allowing the model to distinguish early and late vertices within each chain independently of the global sequence index (Ma et al., 28 Mar 2026).

Inference proceeds by encoding mesh topology and point-cloud semantics, fusing them by cross-attention, and repeatedly generating chains: a start vertex is sampled, the model walks locally under the 1-ring neighbor mask until [EOC] is produced, and generation terminates when [EOS] appears. The paper also describes an optional divide-and-conquer mode that recursively splits the mesh after a seam loop separates the surface into components, improving scalability.

The evaluation uses two large datasets, TexVerse and GarmentCodeData. After filtering, the paper reports about 300K part-level samples and 110K garment samples, split 90/5/5 for train/validation/test. Training uses AdamW for 30 epochs on 8 RTX 4090 GPUs, with up to 2,048 sampled surface points per mesh. The point-cloud encoder is frozen, while the graph encoder and Transformer are trained end to end.

The baselines span production tools and learning-based methods: xatlas, Blender Smart UV Project, OptCuts, Nuvo, and PartUV, together with coordinate-based ablations Coord-Edge and Coord-Chain. Evaluation is intentionally multi-dimensional: overall area distortion, angular distortion, number of charts, island compactness, island convexity, normalized seam length, boundary jaggedness, and a user study are all reported. The paper states that MeshTailor does not always achieve the absolute best distortion, but it produces much better seam structure; it achieves the best boundary jaggedness on both datasets and strong island compactness/convexity. A 2AFC user study with 100 participants and 5,000 votes found MeshTailor consistently preferred over all baselines, with judgments based on minimization, concealment, and geometry awareness (Ma et al., 28 Mar 2026).

4. Relation to prior seam methods, conceptual contribution, and limitations

MeshTailor is positioned against two broad classes of prior methods. First, classical optimization-based approaches, including OptCuts, Autocuts, and related surface parameterization methods, try to balance distortion, cut length, and injectivity. These methods are mathematically principled, but the paper argues that they remain fundamentally local in their objectives: they optimize measurable geometry error without encoding the higher-level structural and semantic logic that human artists use when placing seams. As described in the paper, this can yield technically valid yet visually awkward solutions such as spiral cuts, fragmented islands, or boundaries that ignore symmetry and natural part structure (Ma et al., 28 Mar 2026).

Second, the paper critiques extrinsic learning-based methods. PartUV predicts segmentation in a volumetric or part-tree style representation and then maps the result back onto the mesh, which can create aliasing and misalignment. SeamGPT is characterized as more direct in its autoregressive seam generation, but it predicts seam geometry in Euclidean space and then snaps it to the mesh. MeshTailor identifies two failure modes in this regime: projection artifacts, where small coordinate errors become jagged or misaligned seams after nearest-edge snapping, and fragile snapping heuristics, which can fail on close, folded, or self-occluding geometry such as overlapping trouser legs. MeshTailor’s core claim is that generation should happen intrinsically on the mesh graph, thereby eliminating projection and snapping as separate postprocessing stages.

The conceptual contribution of the method is therefore not merely the use of a Transformer, but a change in problem formulation. Seams are generated as intrinsic walks on the mesh graph; ChainingSeams turns an unordered seam graph into a stable autoregressive target with a coarse-to-fine bias toward global structural cuts; the dual-stream encoder combines connectivity and global geometry; and the pointer layer restricts traversal to valid local neighborhoods. This suggests that the method targets a different notion of quality from classical UV optimization: not only distortion control, but also editability, coherence, and professional usability.

The paper also documents several limitations. The method is mainly trained on meshes up to about 2,000 triangles, though divide-and-conquer inference helps on denser meshes. It struggles on hair-like or highly spiky assets, which are both geometrically difficult and out of distribution, and on extremely low-poly meshes, where too few vertices limit the available seam paths. Because the decoder is autoregressive, error propagation can occur: if one predicted vertex is wrong, the rest of the chain can go off course. The paper notes, however, that seams are represented as separate chains, so a bad chain can be discarded while keeping the others. A stability test under vertex jitter further indicates that seams remain stable for small perturbations but degrade as noise increases (Ma et al., 28 Mar 2026).

5. MeshTailor as anthropometric human mesh estimation

A distinct use of the name appears in human mesh estimation. In that setting, MeshTailor is a pipeline for producing consistent human body shapes across video frames by separating shape from pose and using anthropometric measurements as the bridge between them. The motivating claim is that a person’s basic body shape should stay fixed within a short video, whereas state-of-the-art human mesh estimation models often predict slightly different shapes for each frame, creating visible and physically incorrect inconsistencies. The paper also reports that some existing datasets contain inconsistent ground truth shape annotations, so the supervision itself may be noisy or self-contradictory (Ludwig et al., 2024).

The system uses 36 anthropometric measurements: 23 lengths and 13 circumferences. Examples listed in the paper include height, shoulder width, arm length, forearm length, thigh length, calf length, waist circumference, chest circumference, hip circumference, and head circumference. These measurements are computed on the standard SMPL-X T-pose using predefined landmarks. For lengths, the method uses Euclidean distances or vertical differences; for circumferences, it intersects the mesh with a plane and computes the convex hull of the body contour.

The learned mapping is A2B, from anthropometric measurements to SMPL-X shape parameters:

$\mathcal{F}$2

A2B predicts SMPL-X body-shape parameters, not pose, and separate models are trained for male, female, and neutral subjects. The number of $\mathcal{F}$3 parameters depends on the SMPL-X configuration used in AGORA: 11 for male, 10 for female, and 16 for neutral. Two regressor families are evaluated: SVR with an RBF kernel and small neural networks. The best reported SVR hyperparameters are RBF kernel, $\mathcal{F}$4, and $\mathcal{F}$5; the best neural network has 4 layers, 330 neurons per layer, tanh activation, and Xavier initialization. Training uses mean squared error on the predicted $\mathcal{F}$6 parameters.

The reverse mapping, B2A, is deterministic: starting from $\mathcal{F}$7, the method generates a T-pose mesh and computes the 36 anthropometric measurements from that mesh. The paper evaluates A2B in a cycle-consistency-like way by starting from ground-truth $\mathcal{F}$8, computing measurements with B2A, predicting $\mathcal{F}$9 using A2B, and comparing the recovered $\mathbf{c}=(v_1,\dots,v_T)$0 and measurements.

To combine shape with pose, the pipeline uses inverse kinematics (IK). A 3D human pose estimator such as UU outputs 3D joint coordinates only, whereas a full SMPL-X mesh needs pose parameters $\mathbf{c}=(v_1,\dots,v_T)$1, that is, joint rotations. IK fits SMPL-X to the target 3D joints by minimizing the error between the estimated joints and the joints regressed from the fitted mesh, with a VPoser prior to penalize implausible poses and a $\mathbf{c}=(v_1,\dots,v_T)$2-prior to avoid extreme body shapes. The optimization is described as an iterative gradient-descent-based procedure over both pose $\mathbf{c}=(v_1,\dots,v_T)$3 and shape $\mathbf{c}=(v_1,\dots,v_T)$4. In practice, IK is run frame-by-frame, with each frame initialized from the previous frame’s solution.

The full inference pipeline is: run ViTPose on video frames to obtain 2D keypoints; feed 2D sequences into UU to obtain 3D joint sequences; use IK + VPoser to fit a full SMPL-X pose; obtain a fixed body-shape vector $\mathbf{c}=(v_1,\dots,v_T)$5 from A2B; and combine pose from UU + IK with shape from A2B. The paper evaluates this approach on ASPset and fit3D, reporting MPJPE as the main metric, together with the standard deviation of body height per person, consistency of $\mathbf{c}=(v_1,\dots,v_T)$6, anthropometric reconstruction error, and the fraction of frames with no detection. It states that MPJPE can be lowered by over 30 mm compared to SOTA HME models, and that replacing shape estimates of existing HME models with A2B results both increases performance and guarantees consistent body shapes (Ludwig et al., 2024).

6. Nomenclature, adjacent work, and common confusions

Because the term MeshTailor appears in multiple subfields, nomenclature requires care. In current arXiv usage, one meaning denotes cutting seams via generative mesh traversal on general 3D surfaces, while another denotes anthropometric human mesh estimation with consistent body shape across frames. The two are unrelated at the methodological level: one is a mesh-native generative model for UV seams, the other a measurement-driven human body-shape pipeline.

A common confusion arises with MagicTailor, which is not MeshTailor. "MagicTailor: Component-Controllable Personalization in Text-to-Image Diffusion Models" introduces a framework for component-controllable personalization in text-to-image diffusion models, addressing semantic pollution with Dynamic Masked Degradation and semantic imbalance with Dual-Stream Balancing. The paper explicitly notes that it is not about “MeshTailor”; the closest mesh-related terminology appears only in an applications section, where MagicTailor can be combined with InstantMesh to enable fine-grained design of 3D mesh (Zhou et al., 2024).

A second adjacent line of work is "Stitched Embeddings: A Unified Latent Space for 3D Garments and 2D Patterns", which is directly relevant to mesh tailoring in the garment-manufacturing sense but is not named MeshTailor. That work proposes Stitched Embeddings (StEm), a simulation-free, end-to-end differentiable, bidirectional framework linking 3D garment geometry and 2D sewing pattern parameters in a shared latent space, with BoxMesh as the intermediate representation. It supports pattern reconstruction from meshes and 3D editing from 2D patterns, thereby connecting neural garment reconstruction to manufacturing-oriented pattern space (Sanchietti et al., 1 Jul 2026).

Taken together, these usages show that MeshTailor is not a single stable term across all geometric learning literature. In one branch it names a seam-generation framework whose defining feature is intrinsic traversal of the mesh graph; in another it names a human-mesh pipeline that uses anthropometric measurements to stabilize shape. The broader literature on tailoring-oriented geometry, including MagicTailor and Stitched Embeddings, suggests a wider trend toward methods that preserve production-relevant structure rather than treating meshes solely as generic geometric outputs.