Reproduce “Strips as Tokens: Artist Mesh Generation with Native UV Segmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Rui Xu, Dafei Qin, Kaichun Qiao, Qiujie Dong, Huaijin Pi, Qixuan Zhang, Longwen Zhang, Lan Xu and 3 more
SATO (Strips as Tokens) introduces a generative framework for producing artist-quality 3D meshes by leveraging a novel strip-based tokenization and a multi-stage training pipeline. This approach enables the simultaneous generation of high-fidelity triangle and quadrilateral meshes with native UV segmentation, streamlining 3D asset creation workflows.
Key insights
- Strip-based tokenization with a 'zipper-like' growth strategy effectively preserves edge flow and local surface continuity, leading to more artist-aligned mesh topology and improved compression efficiency.
- Embedding UV island boundaries directly into the token sequence allows the generative model to explicitly predict semantic partitioning alongside geometry, enabling native UV segmentation.
- A unified decoding process, controlled by a vertex stride parameter (δ), allows the same autoregressive model to generate both triangle and quadrilateral meshes without architectural changes, fostering synergistic training.
- A hierarchical training regimen (geometric pretraining followed by UV and quad fine-tuning) is critical for effectively learning complex semantic labels and specialized topologies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.