ChainingSeams in MeshTailor

• ChainingSeams is a hierarchical serialization method that converts unordered UV seam layouts into a single autoregressive vertex-walk sequence, preserving seam connectivity.
• It employs a loops-first, balance-first, and large-patch-first ordering to decompose and prioritize seam chains, enhancing both structural regularity and efficiency.
• Integrated in MeshTailor, the method fuses mesh topology with enriched vertex embeddings via an autoregressive pointer decoder, yielding projection-free, edge-aligned seams.

Searching arXiv for the cited papers and closely related work to ground the article. {"query": "\"MeshTailor\" ChainingSeams arXiv", "max_results": 5} ChainingSeams is a hierarchical serialization of a seam graph on a triangular mesh, introduced in MeshTailor as part of a mesh-native generative framework for synthesizing edge-aligned seams on 3D surfaces (Ma et al., 28 Mar 2026). Its central purpose is to convert an unordered subset of mesh edges representing a UV seam layout into a single autoregressive target sequence while preserving the structural regularities of production seam layouts, in which seams form connected chains and loops rather than independent edge labels. Within MeshTailor, this serialization is coupled to enriched vertex embeddings and an autoregressive pointer decoder so that seams are traced vertex-by-vertex within local neighborhoods, yielding projection-free, edge-aligned seams (Ma et al., 28 Mar 2026).

1. Formal object and graph-theoretic setting

ChainingSeams is defined on a triangular mesh

$\mathcal{M}=(\mathcal{V},\mathcal{E},\mathcal{F}),$

where $\mathcal{V}=\{v_i\}_{i=1}^N$ are the vertices, $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$ are the edges, and $\mathcal{F}$ are the faces (Ma et al., 28 Mar 2026). A UV seam layout is a subset of edges $\mathcal{S}\subseteq\mathcal{E}$, or equivalently a binary labeling of edges,

$$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$

The 1-ring neighborhood of a vertex $v$ is

$\mathcal{N}(v)=\{u\mid (v,u)\in\mathcal{E}\}.$

The key observation underlying ChainingSeams is that seams in production are not merely independent binary labels. They form connected chains and loops on the mesh graph (Ma et al., 28 Mar 2026). This motivates a representation based on maximal vertex walks rather than per-edge classification. In that representation, each connected component of the seam subgraph $(\mathcal{V},\mathcal{S})$ is traced as

$$\mathbf{c}=(v_1,v_2,\dots,v_T),\qquad (v_t,v_{t+1})\in\mathcal{E}\ \forall t.$$

These chains fall into two categories. A loop cut satisfies $\mathcal{V}=\{v_i\}_{i=1}^N$0, whereas an open chain satisfies $\mathcal{V}=\{v_i\}_{i=1}^N$1 (Ma et al., 28 Mar 2026). The total set of chains is denoted by $\mathcal{V}=\{v_i\}_{i=1}^N$2. ChainingSeams then imposes an ordering on $\mathcal{V}=\{v_i\}_{i=1}^N$3 that is not arbitrary, but hierarchical and deterministic.

2. Hierarchical serialization and canonical ordering

The serialization proceeds in two stages: decomposition into seam chains, followed by a coarse-to-fine canonical ordering (Ma et al., 28 Mar 2026). The ordering is explicitly described as:

1. loops-first,
2. balance-first (among loops in the same patch),
3. large-patch-first.

The loops-first rule prioritizes closed cuts before open boundary refinements. The large-patch-first rule maintains a set of active patches $\mathcal{V}=\{v_i\}_{i=1}^N$4, initialized as $\mathcal{V}=\{v_i\}_{i=1}^N$5, and repeatedly selects the current patch of largest area,

$\mathcal{V}=\{v_i\}_{i=1}^N$6

Among loops that lie strictly inside $\mathcal{V}=\{v_i\}_{i=1}^N$7, ChainingSeams evaluates an area-balance score

$\mathcal{V}=\{v_i\}_{i=1}^N$8

where cutting $\mathcal{V}=\{v_i\}_{i=1}^N$9 along loop $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$0 yields two subpatches of areas $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$1 and $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$2 (Ma et al., 28 Mar 2026). The selected loop is

$\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$3

That loop is appended to the sequence, $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$4 is split into its two subpatches, the patch set is updated, and the chosen loop is removed from the loop pool. If no loops remain in the current patch, the patch is discarded. The process repeats until no loop chains remain.

After all loops have been exhausted, the open chains $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$5 are appended, sorted for example by descending chain length, producing a single ordered list

$\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$6

This ordering serves two roles in the MeshTailor formulation. First, it turns an unordered seam graph into a single target sequence for autoregressive learning. Second, it provides what the paper describes as a stable, hierarchical curriculum from global loop cuts down to fine chain details (Ma et al., 28 Mar 2026). The deterministic ordering is therefore not only a serialization device but also a structural prior over seam planning.

3. Integration into MeshTailor

In MeshTailor, ChainingSeams is embedded in a broader architecture comprising a dual-stream encoder, linear sequence construction, and an autoregressive pointer decoder (Ma et al., 28 Mar 2026). The graph stream applies GraphSAGE on $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$7 to obtain per-vertex features

$\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$8

The point-cloud stream samples $\mathcal{E}\subset\mathcal{V}\times\mathcal{V}$9 points on the surface and encodes them with a frozen point-cloud network into tokens

$\mathcal{F}$0

For each vertex $\mathcal{F}$1, cross-attention fuses these two sources: $\mathcal{F}$2 yielding enriched vertex embeddings $\mathcal{F}$3 (Ma et al., 28 Mar 2026).

The ordered chain list is then transformed into a linear token stream,

$\mathcal{F}$4

Here $\mathcal{F}$5 marks the end of a chain and $\mathcal{F}$6 marks the end of the full seam layout.

The decoder is an autoregressive pointer mechanism. At time $\mathcal{F}$7, it produces a hidden state $\mathcal{F}$8 and points into the candidate set of mesh vertices together with the two special tokens $\mathcal{F}$9. The next-token distribution is

$\mathcal{S}\subseteq\mathcal{E}$0

where the mask $\mathcal{S}\subseteq\mathcal{E}$1 enforces mesh-native validity (Ma et al., 28 Mar 2026). If the last token is a vertex $\mathcal{S}\subseteq\mathcal{E}$2, only that vertex’s 1-ring neighbors plus $\mathcal{S}\subseteq\mathcal{E}$3 and $\mathcal{S}\subseteq\mathcal{E}$4 receive zero mask value; all others receive $\mathcal{S}\subseteq\mathcal{E}$5. If the last token is $\mathcal{S}\subseteq\mathcal{E}$6, any vertex may start the next chain.

Two positional systems are added to decoder inputs: global sequence positions using rotary embeddings, and chain-local positions that reset to zero after each $\mathcal{S}\subseteq\mathcal{E}$7 (Ma et al., 28 Mar 2026). The former encodes position in the overall serialization, whereas the latter signals position within an individual chain. Together with the ChainingSeams ordering, these mechanisms let the model distinguish global cut planning from local path continuation.

4. Generative implications of the mesh-native formulation

ChainingSeams is inseparable from MeshTailor’s claim to be the first mesh-native generative framework for synthesizing edge-aligned seams on 3D surfaces (Ma et al., 28 Mar 2026). The framework contrasts with prior optimization-based or extrinsic learning-based methods by operating directly on the mesh graph, thereby eliminating projection artifacts and fragile snapping heuristics.

A useful comparison is provided by SeamCrafter, which also performs autoregressive seam generation but parameterizes seams as sequences of 3D segment endpoints quantized into discrete bins and, after decoding, projects predicted segments onto the mesh topology by nearest-neighbor endpoint assignment followed by shortest geodesic path computation on the mesh graph (Xu et al., 25 Sep 2025). ChainingSeams adopts a different representational commitment: the generative process itself is already expressed as a vertex-walk on the mesh graph (Ma et al., 28 Mar 2026). This mesh-native formulation is what makes the pointer mask naturally local and validity-preserving.

The same contrast appears at the encoder level. SeamCrafter uses a dual-branch point-cloud encoder that disentangles topology and geometry through separate point sets and VecSet-style encoders, with the two streams concatenated before cross-attention into an autoregressive seam decoder (Xu et al., 25 Sep 2025). MeshTailor instead fuses topological and geometric context through a graph stream on the mesh itself and a frozen point-cloud stream, followed by cross-attention into enriched vertex embeddings (Ma et al., 28 Mar 2026). This suggests that ChainingSeams is not only a sequence format but also part of a broader modeling choice in which seam synthesis is treated as traversal over intrinsic mesh adjacency rather than as coordinate prediction followed by topological correction.

5. Empirical validation and ablation evidence

The quantitative validation reported for ChainingSeams is tied to the broader MeshTailor evaluation on TexVerse and GarmentCodeData (Ma et al., 28 Mar 2026). On these two large datasets, MeshTailor achieves among the best trade-offs between distortion and chart usability. Its “boundary jaggedness,” defined as mean discrete curvature

$\mathcal{S}\subseteq\mathcal{E}$8

on UV island outlines, is reported as dramatically lower than all baselines, with the example $\mathcal{S}\subseteq\mathcal{E}$9 versus $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$0 for Smart-UV or OptCuts (Ma et al., 28 Mar 2026). Chart count is also reduced, with an example of approximately $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$1 islands versus $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$2 for xatlas, while keeping compactness and convexity near the professional-layout ground truth.

The ablation findings directly isolate the contribution of the representation. Removing the ChainingSeams ordering, or replacing the mesh-native pointer with coordinate-based generation, raises jaggedness by $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$3–$$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$4 and increases sequence lengths by $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$5–$$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$6 (Ma et al., 28 Mar 2026). These ablations indicate that the hierarchical serialization is not merely an implementation detail; it changes both geometric regularity and generative efficiency.

The evaluation also includes a user study. In pairwise 2AFC tests over $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$7 votes, MeshTailor’s seams are preferred more than $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$8 of the time over each baseline (Ma et al., 28 Mar 2026). The paper attributes this outcome to the coarse-to-fine, loops-first serialization, which yields more coherent, professional-quality seams.

Related work provides a complementary perspective on seam structure. In garment manipulation, a Seam-to-Graph network chains partial 3D seam observations into a topology-encoded structural skeleton graph through point-cloud and seam-segment encoding, cross-attention grounding, GAT-based topology-aware message passing, and post-hoc optimization (Huang et al., 13 Jun 2026). Although this is a different task, it underscores the broader methodological theme that seams are structurally meaningful graph objects rather than isolated local features.

6. Scope, related usages, and conceptual boundaries

Within geometry processing and generative UV layout, ChainingSeams denotes the hierarchical serialization introduced in MeshTailor (Ma et al., 28 Mar 2026). The term “chaining seams,” however, also appears in an unrelated setting in integrability and planar $$y_e = \begin{cases} 1 & e\in\mathcal{S},\ 0 & e\notin\mathcal{S}. \end{cases}$$9 SYM. There, the “Chaining Seams” or “Tailoring” procedure of Escobedo, Gromov, Sever, and Vieira computes tree-level three-point structure constants by cutting three closed spin chains into open subchains, flipping right subchains into bras, and sewing adjacent pieces through spin-chain scalar products of Slavnov type, normalized by Gaudin norms (Escobedo et al., 2010).

That older usage is formal and algebraic rather than geometric. Its basic workflow is “Cut–Flip–Sew,” and its seam language refers to how subchains are assembled into the pair-of-pants contraction pattern of a three-point function (Escobedo et al., 2010). By contrast, ChainingSeams in MeshTailor concerns UV seam graphs on triangular meshes, with loops, open chains, active patches, 1-ring neighborhoods, and pointer-based traversal (Ma et al., 28 Mar 2026). The shared vocabulary therefore reflects a generic notion of ordered assembly rather than a shared application domain.

A plausible implication is that the term has become useful whenever an initially unordered seam-like structure must be made sequential and compositional. In MeshTailor, that means converting a seam graph into an ordered chain list for autoregressive learning (Ma et al., 28 Mar 2026). In garment configuration alignment, it means grounding partial seam segments into a coherent skeleton graph (Huang et al., 13 Jun 2026). In integrability, it means assembling subchain overlaps into a three-point observable (Escobedo et al., 2010). The commonality is procedural serialization; the mathematical objects and objectives are otherwise distinct.

7. Significance for autoregressive seam generation

ChainingSeams occupies a specific methodological position in seam synthesis. It addresses the mismatch between unordered graph targets and sequence models by encoding seam layouts as vertex-walks ordered from globally consequential loop cuts to local open-chain details (Ma et al., 28 Mar 2026). This representation is compatible with local validity masks, mesh-native pointer decoding, and vertex-level embeddings enriched by both topology and geometry.

Its significance is clearest when contrasted with extrinsic coordinate-generation approaches. Coordinate-based segment generation can require post hoc projection onto mesh topology, while ChainingSeams expresses the target directly in the combinatorial language of the mesh graph (Xu et al., 25 Sep 2025). The resulting sequence is not only valid by construction under the decoder mask, but also organized to expose long-range seam structure early in the decoding process (Ma et al., 28 Mar 2026).

The representation therefore functions simultaneously as a serialization scheme, a training curriculum, and a structural prior over seam planning. In the MeshTailor formulation, these three roles are inseparable: the hierarchy determines the token order, the token order shapes the autoregressive learning problem, and the learning problem benefits from the fact that global loop cuts are resolved before fine chain details (Ma et al., 28 Mar 2026).