---
title: StructRe
canonical_url: "https://www.modelscope.cn/models/wangjiepeng/StructRe"
md_url: "https://www.modelscope.cn/models/wangjiepeng/StructRe.md"
repository: wangjiepeng/StructRe
last_updated: 2026-08-09
license: mit
downloads: 0
stars: 1
tags:
  - 3d-generation
  - shape-generation
  - 3d-reconstruction
  - mesh-generation
  - point-cloud
---

# StructRe

> StructRe - wangjiepeng 在 ModelScope 开源的模型。StructRe: Rewriting for Structured Shape Modeling

wangjiepeng/StructRe 是 ModelScope 魔搭社区上的机器学习模型，采用 mit 许可。

- **Repository**: wangjiepeng/StructRe
- **License**: mit
- **Tags**: 3d-generation, shape-generation, 3d-reconstruction, mesh-generation, point-cloud
- **Downloads**: 0
- **Stars**: 1
- **Last updated**: 2026-08-09

Source: https://www.modelscope.cn/models/wangjiepeng/StructRe

---

# StructRe: Rewriting for Structured Shape Modeling

StructRe is a three-stage framework for structured 3D shape modeling. It uses
local and probabilistic structure rewriting to provide a unified solution for
structured shape reconstruction, completion, and generation.

[Project page](https://jiepengwang.github.io/StructRe/) |
[Paper](https://arxiv.org/abs/2311.17510) |
[Code](https://github.com/jiepengwang/StructRe/)

This repository hosts the pretrained model checkpoints. Please use them with
the official StructRe codebase linked above.

## Model architecture

StructRe consists of three cooperating stages:

| Stage | Model | Purpose |
| --- | --- | --- |
| 1 | SparseVAE | Encodes and decodes local part geometry with sparse 3D convolutions |
| 2 | PartVAE | Represents part arrangements in a structural latent space |
| 3 | StructMapping | Performs iterative structure rewriting and produces hierarchical shapes through the Stage 1 and Stage 2 decoders |

Stage 3 inference requires all three checkpoints.

## Checkpoints

| File | Model | Size |
| --- | --- | ---: |
| [`ckpt/sparse_vae/ckpt_0048.pt`](./ckpt/sparse_vae/ckpt_0048.pt) | SparseVAE | 146 MB |
| [`ckpt/part_vae/ckpt_0046.pt`](./ckpt/part_vae/ckpt_0046.pt) | PartVAE | 798 MB |
| [`ckpt/structmapping/ckpt_0005.pt`](./ckpt/structmapping/ckpt_0005.pt) | StructMapping | 862 MB |

## Quick start

Clone the source code and this checkpoint repository:

```bash
git clone https://github.com/jiepengwang/StructRe.git
git clone https://www.modelscope.cn/wangjiepeng/StructRe.git
cd StructRe
```

Point the codebase to the downloaded checkpoints:

```bash
export STRUCTRE_STAGE1_CHECKPOINT=/path/to/StructRe-checkpoints/ckpt/sparse_vae/ckpt_0048.pt
export STRUCTRE_STAGE2_CHECKPOINT=/path/to/StructRe-checkpoints/ckpt/part_vae/ckpt_0046.pt
export STRUCTRE_STAGE3_CHECKPOINT=/path/to/StructRe-checkpoints/ckpt/structmapping/ckpt_0005.pt
```

The reference environment uses Python 3.12, PyTorch 2.9.1 with CUDA 12.8,
MinkowskiEngine 0.5.4, and PyTorch3D 0.7.9. Installation, data preparation,
training, and inference instructions are available in the
[code repository](https://github.com/jiepengwang/StructRe/).


## Citation

If you use StructRe in your research, please cite:

```bibtex
@article{wang2023structre,
  title   = {StructRe: Rewriting for Structured Shape Modeling},
  author  = {Wang, Jiepeng and Pan, Hao and Liu, Yang and Tong, Xin and Komura, Taku and Wang, Wenping},
  journal = {arXiv preprint arXiv:2311.17510},
  year    = {2023}
}
```

## Acknowledgments

We gratefully acknowledge the authors and contributors of [AutoSDF](https://github.com/yccyenchicheng/AutoSDF), [MinkowskiEngine](https://github.com/NVIDIA/MinkowskiEngine), [DETR](https://github.com/facebookresearch/detr), [StructureNet](https://github.com/daerduoCarey/structurenet), and [BlenderToolBox](https://github.com/HTDerekLiu/BlenderToolbox) for making their code and tools publicly available.

## License

StructRe is released under the [MIT License](https://github.com/jiepengwang/StructRe/blob/main/LICENSE).