Reproduce “MeshMosaic: Scaling Artist Mesh Generation via Local-to-Global Assembly” 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, Tianyang Xue, Qiujie Dong, Le Wan, Zhe Zhu, Peng Li, Zhiyang Dou, Cheng Lin and 4 more
MeshMosaic turns input point clouds into detailed triangular meshes by generating locally quantized patches conditioned on the full shape and previously generated neighboring boundaries. On 100 Thingi10K test shapes, its edge F-score was 0.271 versus 0.179 for the strongest baseline. Distant parts can remain weakly coupled, and meshes above 100K faces may take hours.
Key insights
- Decomposing global mesh generation into local, boundary-conditioned patch generation effectively circumvents the long-sequence bottleneck of autoregressive transformers.
- Applying local quantization to individual patches dramatically boosts the effective resolution and detail preservation, allowing for faithful reconstruction of intricate features.
- Explicitly conditioning the autoregressive model on both local patch features, global shape features, and GRU-encoded boundary information is crucial for maintaining seamless connectivity and overall coherence.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.