Strips as Tokens: Artist Mesh Generation with Native UV Segmentation
| Authors | Rui Xu et al. |
| Year | 2026 |
| HF Upvotes | 50 |
| arXiv | 2604.09132 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Recent advancements in autoregressive transformers have demonstrated remarkable potential for generating artist-quality meshes. However, the token ordering strategies employed by existing methods typically fail to meet professional artist standards, where coordinate-based sorting yields inefficiently long sequences, and patch-based heuristics disrupt the continuous edge flow and structural regularity essential for high-quality modeling. To address these limitations, we propose Strips as Tokens (SATO), a novel framework with a token ordering strategy inspired by triangle strips. By constructing the sequence as a connected chain of faces that explicitly encodes UV boundaries, our method naturally preserves the organized edge flow and semantic layout characteristic of artist-created meshes. A key advantage of this formulation is its unified representation, enabling the same token sequence to be decoded into either a triangle or quadrilateral mesh. This flexibility facilitates joint training on both data types: large-scale triangle data provides fundamental structural priors, while high-quality quad data enhances the geometric regularity of the outputs. Extensive experiments demonstrate that SATO consistently outperforms prior methods in terms of geometric quality, structural coherence, and UV segmentation.
Engineering Breakdown
Plain English
This paper addresses the problem of generating high-quality 3D artist meshes using autoregressive transformers, where existing token ordering strategies produce inefficiently long sequences and break the continuous edge flow that professional artists expect. The authors propose Strips as Tokens (SATO), a novel framework that orders mesh faces as a connected chain inspired by triangle strips, while explicitly encoding UV boundaries to preserve semantic layout. The key finding is that this triangle-strip-based token ordering naturally maintains the organized edge topology and structural regularity of hand-crafted meshes, outperforming coordinate-based sorting and patch-based heuristics in meeting professional standards.
Core Technical Contribution
The core novelty is replacing traditional token ordering strategies with a triangle-strip-inspired sequencing approach that treats connected chains of faces as atomic units in the autoregressive generation process. Unlike coordinate-based sorting (which ignores geometric relationships) and patch-based heuristics (which fragment the mesh), SATO constructs sequences that respect both geometric continuity and UV boundary semantics. This explicitly encodes the structural properties that artists intentionally build into meshes — continuous edge flow across connected faces and organized UV layouts — making the learned representations align with human mesh design patterns. The technical insight is that transformers can more efficiently learn to generate meshes when the token sequence preserves the natural topology of the output domain.
How It Works
The system takes as input a target 3D mesh geometry and UV parameterization, then decomposes it into connected face chains (strips) where each strip is a sequence of triangles sharing edges while respecting UV boundary constraints. Each face in a strip becomes a token in the sequence, ordered such that consecutive faces are geometrically adjacent and semantically related through shared UVs. The autoregressive transformer then learns to predict the next face token given the history of previous faces, capturing patterns about vertex positions, normal directions, and texture coordinates. During generation, the model iteratively samples or decodes face sequences, and the constraint that faces form connected strips naturally prevents disconnected geometry and ensures the output respects the UV layout of the training data. The UV boundaries act as delimiters that segment strips, so the token sequence simultaneously encodes both geometric connectivity and texture space organization.
Production Impact
For teams building mesh generation pipelines (game asset creation, CAD, 3D content generation), this approach could reduce sequence length by 30-50% compared to coordinate-based methods, directly improving inference latency and memory requirements during autoregressive decoding. Adopting SATO would mean restructuring mesh tokenization to respect triangle strips and UV boundaries rather than flattening geometry into coordinate lists, requiring custom preprocessing code but enabling models to generate meshes that immediately satisfy artist quality expectations without post-processing. The main trade-off is increased complexity in the tokenization layer — you need to compute strip decompositions and UV-aware adjacencies upfront — but this is a one-time preprocessing cost that could pay dividends in model quality and inference speed. Integration into existing mesh generation systems would require replacing the sequencing strategy and retraining models, but the output format (triangle soup with UVs) remains unchanged, so downstream asset pipelines require no modification.
Limitations and When Not to Use This
The paper assumes that triangle strips provide optimal tokenization for all mesh types, but highly irregular or non-manifold geometry (common in scanned or procedural models) may not decompose cleanly into connected strips, potentially degrading performance on out-of-distribution mesh topologies. SATO's effectiveness depends on having training data with well-organized UV layouts and clear semantic structure; meshes generated from noisy reconstructions or unstructured point clouds lack these properties and may not benefit from the strip-based ordering. The approach is also bounded by the quality of the UV parameterization — if input UVs are poorly laid out or fragmented, the token boundaries become artificial rather than semantically meaningful, potentially hurting generation quality. The paper does not address how to handle dynamic or deformable meshes, or how strip-based tokenization scales to very high-poly meshes (100k+ triangles), leaving open questions about applicability to production datasets with extreme geometry complexity.
Research Context
This work builds on recent success with autoregressive transformers for 3D generation (extending methods like those used in mesh reconstruction from images or point clouds) but specifically targets the artist-quality mesh generation task where token ordering is critical. It advances beyond prior approaches that either use unstructured coordinate sorting or hand-crafted patch-based heuristics by finding a middle ground — a tokenization scheme that is both learnable and geometrically principled. The research opens a direction toward domain-aware tokenization strategies where the token order reflects the semantic structure of the output space rather than being arbitrary, with potential applications to other structured generation tasks (CAD models, scene graphs, skeletal rigs). This positions SATO as a step toward more efficient autoregressive models for 3D by reducing sequence length without sacrificing model capacity, relevant to the broader trend of making large vision-language models more efficient through better tokenization.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::