[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-80803":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":9,"htmlUrl":10,"language":11,"languages":10,"totalLinesOfCode":10,"stars":12,"forks":13,"watchers":14,"openIssues":15,"contributorsCount":13,"subscribersCount":13,"size":13,"stars1d":16,"stars7d":17,"stars30d":18,"stars90d":13,"forks30d":13,"starsTrendScore":17,"compositeScore":13,"rankGlobal":10,"rankLanguage":10,"license":19,"archived":20,"fork":20,"defaultBranch":21,"hasWiki":22,"hasPages":20,"topics":23,"createdAt":10,"pushedAt":10,"updatedAt":28,"readmeContent":29,"aiSummary":30,"trendingCount":13,"starSnapshotCount":13,"syncStatus":16,"lastSyncTime":31,"discoverSource":32},80803,"LATO","TianhaoZhao668\u002FLATO","TianhaoZhao668","Offical repo for the paper \"LATO: 3D Mesh Flow Matching with Structured TOpology Preserving LAtents\"(ICML'26)","https:\u002F\u002Ftianhaozhao668.github.io\u002FLATO\u002F",null,"Python",46,0,4,1,2,6,8,"MIT License",false,"main",true,[24,25,26,27],"3d","3d-aigc","3d-generation","3d-reconstruction","2026-06-12 02:04:07","# LATO: 3D Mesh Flow Matching with Structured TOpology Preserving LAtents\n\n\u003Cdiv align=\"center\">\n\n[![🏠 Project Page](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FProject-Page-blue)](https:\u002F\u002Ftianhaozhao668.github.io\u002FLATO\u002F)\n[![📄 Paper](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FPaper-arXiv-green)](https:\u002F\u002Farxiv.org\u002Fpdf\u002F2603.06357)\n[![🤗 Model](https:\u002F\u002Fimg.shields.io\u002Fbadge\u002FModel-Hugging%20Face-yellow)](https:\u002F\u002Fhuggingface.co\u002Fudbbdh\u002FLATO)\n\n\u003C\u002Fdiv>\n\n![teaser](assert\u002Fdocs\u002Fteaser.png)\n\n\n## 🌟 Overview\nLATO introduces a novel topology-preserving latent representation by natively compressing mesh topology into a structured latent space. Unlike traditional implicit methods that rely on post-hoc isosurface extraction (e.g., Marching Cubes) and autoregressive model, LATO ensures high-fidelity generation of explicit meshes with effieient topological connectivity.\n\nBy modeling the mesh as a Vertex Displacement Field (VDF) anchored on the surface geometry, we successfully map dense explicit signals into a differentiable, topology-aware latent space. Our framework enables:\n- 🧩 Topology-aware mesh representation\n- 🔗 Explicitly decodes connectivity\n- 📉 Memory-Efficient training bypasses $O(N^2)$ complexity\n- ⚡ Generates artistic meshes in seconds\n\n\n## 🔥 News\n\n[2026-05-14] Initial Release:\n  - Pretrained VAE model weights ($512^3$ reconstruction)\n  - Inference scripts and examples\n  - LATO implementation\n\n[2026-05-01] 🎉🎉 Our paper was received by ICML2026.\n\n\n## 🚀 Quick Start\n\n### Installation\n\n1. Clone the repository:\n\n```bash\ngit clone https:\u002F\u002Fgithub.com\u002FTianhaoZhao668\u002FLATO.git\ncd LATO\n```\n\n2. Install dependencies:\n\n```bash\n# 1. Create conda environment\nconda create -n lato python=3.10.20 -y\nconda activate lato\npython -m pip install --upgrade pip setuptools wheel packaging ninja\n\n# 2. Install system libraries required by open3d :\nsudo apt-get update\nsudo apt-get install -y libgl1 libglib2.0-0\n\n# 3. Install the pinned Python dependencies:\npip install -r requirements.txt\n\n# 4. Install flash-attn wheel and install:\nwget https:\u002F\u002Fgithub.com\u002FDao-AILab\u002Fflash-attention\u002Freleases\u002Fdownload\u002Fv2.6.3\u002Fflash_attn-2.6.3+cu118torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl\npip install flash_attn-2.6.3+cu118torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl \n```\n\n\n### Download Checkpoint\n\nDownload the released LATO 512 VAE checkpoint from [Hugging Face](https:\u002F\u002Fhuggingface.co\u002FYOUR_ORG\u002FLATO) or:\n\n```bash\nhf download udbbdh\u002FLATO \\\n  checkpoints\u002F128to512\u002Fvae\u002Fvae_128to512.pt \\\n  --local-dir . \n```\n\n\n### Running Inference\nBasic reconstruction can be:\n```bash\npython scripts\u002Finfer_vae_512.py \\\n  --mesh assert\u002Fsample\u002Ftest.obj \\\n  --checkpoint checkpoints\u002F128to512\u002Fvae\u002Fvae_128to512.pt \\\n  --config configs\u002Finfer_vae_512.yaml \\\n  --output outputs\n```\n\n\n## 📊 Technical Details\nLATO Architecture & Pipeline:\n\n- **Input**: Active voxels, point clouds with position, normal and VDF\n- **Encoder**: Sparse transformer for efficient topologt compression\n- **Decoder**: Decoding vertex and edge using prune head and connection head\n- **Output**: Topological 3D meshes\n\n\n## 🙏 Acknowledgements\n\nOur work builds upon these excellent repositories:\n\n- [TripoSF](https:\u002F\u002Fgithub.com\u002FVAST-AI-Research\u002FTripoSF) \n- [TRELLIS](https:\u002F\u002Fgithub.com\u002FMicrosoft\u002FTRELLIS)\n\n\n\n## 📝 Citation\n\nIf you find LATO useful in your research, please consider citing:\n\n```bibtex\n@article{zhao2026lato,\n  title={LATO: 3D Mesh Flow Matching with Structured TOpology Preserving LAtents},\n  author={Zhao, Tianhao and Zhang, Youjia and Long, Hang and Zhang, Jinshen and Li, Wenbing and Yang, Yang and Zhang, Gongbo and Hladk{\\`y}, Jozef and Nie{\\ss}ner, Matthias and Yang, Wei},\n  journal={arXiv preprint arXiv:2603.06357},\n  year={2026}\n}\n```\n\n","LATO 是一个用于3D网格流匹配的项目，通过将网格拓扑结构压缩到结构化的潜在空间中，实现了高保真的显式网格生成。其核心功能包括基于顶点位移场（VDF）的拓扑感知网格表示、显式解码连通性以及高效的内存使用训练方法，绕过了传统方法中的O(N^2)复杂度问题。LATO 采用稀疏变换器作为编码器，并通过剪枝头和连接解码器来生成网格的顶点和边。该项目特别适用于需要高效生成具有复杂拓扑结构的3D模型的应用场景，如艺术设计、虚拟现实和游戏开发等。","2026-06-11 04:02:23","CREATED_QUERY"]