DI-PCG: Diffusion-based Efficient Inverse Procedural Content Generation for High-quality 3D Asset Creation

   

Wang Zhao¹, Yan-Pei Cao², Jiale Xu¹, Yuejiang Dong^1,3, Ying Shan¹

¹ARC Lab, Tencent PCG   ²VAST   ³Tsinghua University

---

🚩 Overview

This repository contains code release for our technical report "DI-PCG: Diffusion-based Efficient Inverse Procedural Content Generation for High-quality 3D Asset Creation".

⚙️ Installation

First clone this repository:

git clone https://github.com/TencentARC/DI-PCG.git
cd DI-PCG

We recommend using anaconda to install the dependencies:

conda create -n di-pcg python=3.10.14
conda activate di-pcg
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0  pytorch-cuda=11.8 -c pytorch -c nvidia
pip install -r requirements.txt

🚀 Usage

For a quick start, try the huggingface gradio demo here.

Download models

We provide the pretrained diffusion models for chair, vase, table, basket, flower and dandelion. You can download them from model card and put them in ./pretrained_models/.

Alternatively, the inference script will automatically download the pretrained models for you.

Local gradio demo

To run the gradio demo locally, run:

python app.py

Inference

To run the inference demo, simply use:

python ./scripts/sample_diffusion.py --config ./configs/demo/chair_demo.yaml

This script processes all the chair images in the ./examples/chair folder and saves the generated 3D models and their rendered images in ./logs.

To generate other categories, use the corresponding YAML config file such as vase_demo.yaml. Currently we supprt chair, table, vase, basket, flower and dandelion generators developped by Infinigen.

python ./scripts/sample_diffusion.py --config ./configs/demo/vase_demo.yaml

Training

We train a diffusion model for each procedural generator. The training data is generated by randomly sampling the PCG and render multi-view images. To prepare the training data, run:

python ./scripts/prepare_data.py --generator ChairFactory --save_root /path/to/save/training/data

Replace ChairFactory with other category options as detailed in the ./scripts/prepare_data.py file. This script also conducts offline augmentation and saves the extracted DINOv2 features for each image, which may consume a lot of disk storage. You can adjust the number of the generated data and the render configurations accordingly.

After generating the training data, start the training by:

python ./scripts/train_diffusion.py --config ./configs/train/chair_train.yaml

Use your own PCG

DI-PCG is general for any procedural generator. To train a diffusion model for your PCG, you need to implement the get_params_dict, update_params, spawn_assets, finalize_assets functions and place your PCG in ./core/assets/. Also change the num_params in your training YAML config file.

If you have any question, feel free to open an issue or contact us.

📚 Citation

If you find our work useful for your research or applications, please cite using this BibTeX:

@article{zhao2024dipcg,
  title={DI-PCG: Diffusion-based Efficient Inverse Procedural Content Generation for High-quality 3D Asset Creation},
  author={Zhao, Wang and Cao, Yanpei and Xu, Jiale and Dong, Yuejiang and Shan, Ying},
  journal={arXiv preprint arXiv:2412.15200},
  year={2024}
}

🤗 Acknowledgements

DI-PCG is built on top of some awesome open-source projects: Infinigen, Fast-DiT. We sincerely thank them all.