VertexRegen: Mesh Generation with Continuous Level of Detail

Unofficial implementation of the ICCV 2025 paper VertexRegen: Mesh Generation with Continuous Level of Detail.

---

📖 Overview

VertexRegen is a mesh generation framework that supports continuous levels of detail (LoD), enabling smooth transitions across mesh resolutions. This repository provides an end-to-end re-implementation of the original paper, including:

• Dataset preprocessing and vertex-split generation
• Mesh tokenization
• Model training
• Distributed inference

The goal of this project is to offer a clean, reproducible research codebase and practical tooling for experimenting with continuous-LoD mesh generation.

---

🚧 Project Status

• Data generation & vertex-split preprocessing
• Tokenization pipeline
• Training & inference code
• Pre-trained weights on Objaverse / Objaverse-XL

⚠️ Note: Only overfitting configurations are currently released to validate the training and inference pipelines.

---

🧩 Tokenization Library

A standalone VertexRegen tokenizer is available on PyPI:

👉 https://pypi.org/project/vertexregen-tokenizer/

You can either install it directly from PyPI or build it from source.

Install from source

Requires libcgal-dev and related CGAL system dependencies.

pip install -e tokenizer/

---

🛠️ Installation

Python environment

Please ensure you have PyTorch ≥ 2.9.1, then install the remaining dependencies:

pip install -r requirements.txt

Docker (recommended)

A pre-built Docker image is provided for convenience:

docker pull zx1239856/vertexregen:0.1.0

---

📦 Data Preparation

You can generate vertex-split datasets using the provided demo ShapeNet data (converted from MeshGPT format). Other datasets are also supported as long as they contain the following fields:

• uid
• vertices
• faces

1️⃣ Generate vertex-split data

python -m scripts.create_dataset -o dataset/collapsed_shapenet_q128 -q 128

2️⃣ Run demo tokenization

python -m scripts.demo_tokenize -i dataset/collapsed_shapenet_q128/ -o demo

---

🧠 Training & Inference

Training

python launch.py --num-processes=4 train.py --config-name=shapenet_uncond_opt_350_overfit

Inference (distributed)

Replace the model path with your actual checkpoint directory. Use -h for additional options.

python launch.py --num-processes=4 infer.py --model outputs/.../checkpoint-500/

ℹ️ Reproducibility note: Fully reproducing the results from the paper requires training on large-scale datasets such as Objaverse-XL, which are not included in this release.

---

🏷️ License

This repository is released under the MIT license.

🪶 Citation

If you find this repository useful, please cite the original paper:

@InProceedings{Zhang_2025_ICCV_VertexRegen,
    author    = {Zhang, Xiang and Siddiqui, Yawar and Avetisyan, Armen and Xie, Chris and Engel, Jakob and Howard-Jenkins, Henry},
    title     = {VertexRegen: Mesh Generation with Continuous Level of Detail},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {12570--12580}
}