microsoft/SQuadGen

SQuadGen: Generating Simple Quad Layouts via Chart Distance Fields

Introduction

SQuadGen is a diffusion-based generative framework for synthesizing simple quad layouts on 3D shapes. It introduces Chart Distance Fields (CDF), a continuous surface-based representation that makes quad-layout generation amenable to neural learning while preserving structure useful for downstream modeling and editing.

This repository hosts the pretrained checkpoints and matching configuration files used by the official inference scripts. It is intended as the model release page; installation, data preparation, training, evaluation, and full inference instructions are maintained in the GitHub repository.

Model Details

• Developed by: Youkang Kong, Yang Liu, Yue Dong, Xin Tong, Heung-Yeung Shum
• Affiliation: Tsinghua University, Microsoft Research Asia
• Model Type: Latent diffusion transformer (SiT-XL) conditioned on Geometry-AE representations and decoded with SQ-VAE
• Parameters: Geom-AE approximately 107M; SQ-VAE approximately 112M; SQ-Diffuse approximately 802M
• Input: Triangle mesh
• Output: Per-vertex CDF / DCDF fields, from which simple quad layouts are extracted

Files

ckpts/
sqvae_geomae.pth # Combined SQ-VAE (~112M) and Geom-AE (~107M) checkpoint
sqvae_geomae.yaml # matching config
sqdiffuse.pth # SQ-Diffuse diffusion checkpoint (~802M)
sqdiffuse.yaml # matching config

sqvae_geomae.pth contains both the SQ-VAE and the Geom-AE. SQ-Diffuse inference requires loading both sqvae_geomae.pth and sqdiffuse.pth.

Download

Download the complete model repository, including the checkpoints and configuration files:

hf download microsoft/SQuadGen --local-dir ./ckpts_hf

The released files are native PyTorch checkpoints. The sqvae_geomae.pth checkpoint contains both the SQ-VAE and Geometry-AE components; SQ-Diffuse inference requires both this checkpoint and sqdiffuse.pth.

Usage

All inference, training and evaluation scripts live in the official GitHub repository. After cloning and setting up the environment as described there, point the --ae_pth / --model_pth arguments at the corresponding files downloaded from this repo, e.g.:

After installing the environment described in the GitHub repository, pass ./ckpts_hf/ckpts/sqvae_geomae.pth and ./ckpts_hf/ckpts/sqdiffuse.pth to the corresponding inference scripts.

Inference Speed (NVIDIA A100 GPU)

| Configuration | 4096 points | 8192 points | | ------------------------- | ----------: | -----------: | | Geom-AE encoding | 0.4s | 1.2s | | SQ-VAE encoding*/decoding | 0.4s/0.4s | 1.3s/1.2s | | SQ-Diffuse generation | 65.9s | 171.7s |

*include Geom-AE encoding

Data and Evaluation

The models were trained on publicly available 3D geometry datasets, including Objaverse, Amazon Berkeley Objects (ABO), and ShapeNet. Dataset-specific terms, attribution requirements, and permitted uses remain applicable to users who work with those datasets or derived assets.

Performance was evaluated on academic quad-meshing and reconstruction benchmarks described in the paper. Robustness coverage is limited to the tested datasets; the release does not claim a unified standardized comparison against all third-party models.

Responsible AI Considerations

SQuadGen is a research prototype for 3D geometry encoding and quad-layout generation. It is not intended for safety-critical, high-risk, real-time, or fully autonomous use, including production CAD, engineering, simulation, or manufacturing workflows, unless additional validation and safeguards are applied.

The models are trained on publicly available 3D datasets, including Objaverse, ABO, and ShapeNet. Their performance may therefore reflect gaps in dataset coverage: results are generally stronger for common object categories and clean meshes, and may degrade for rare or highly complex shapes, noisy or non-manifold meshes, irregular topology, strict CAD constraints, or other out-of-distribution inputs. The models do not provide guarantees of geometric correctness, mesh validity, manufacturability, robustness, or consistency across datasets and mesh resolutions.

Generated layouts and intermediate representations should be treated as learned approximations. Before downstream use, developers should validate input geometry and output quality, use fallback heuristics where appropriate, and retain human review for decisions that affect design, simulation, engineering, or manufacturing. Users should also verify that input geometry is legally and appropriately sourced. These models do not process language or make decisions in sensitive domains such as healthcare or legal services.

Citation

@article{kong2026squadgen,
author = {Youkang Kong and Yang Liu and Yue Dong and Xin Tong and Heung-Yeung Shum},
title = {{SQuadGen}: Generating Simple Quad Layouts via Chart Distance Fields},
journal = {ACM Transactions on Graphics (SIGGRAPH)},
volume = {45},
number = {4},
pages = {144:1--144:15},
year = {2026},
doi = {10.1145/3811348}
}

License

This model is released under the MIT License.