PhysGen

📄 Paper · 📚 arXiv · 💻 Code

PhysGen: Physically Grounded 3D Shape Generation for Industrial Design
Accepted to CVPR 2026

---

Overview

PhysGen is a physics-grounded 3D shape generation framework for industrial design. It jointly models geometry and physical properties in a shared latent space and enables physics-guided generation for both unconditional and conditional settings.

This repository includes:

• training and evaluation code for the Shape-Physics Variational Autoencoder (SP-VAE)
• training and inference code for physics-guided flow matching model
• preprocessing scripts for the DrivAerNet++ dataset
• pretrained checkpoints for reproduction

---

Preparation

1. Environment Setup

This project is developed and tested on NVIDIA A100/H100 GPUs. We recommend using an Anaconda environment.

git clone https://github.com/kasvii/PhysGen.git
cd PhysGen

conda create -n physgen python=3.10 -y
conda activate physgen
conda install -y pytorch=2.4.0 torchvision=0.19.0 torchaudio=2.4.0 pytorch-cuda=12.1 -c pytorch -c nvidia
pip install -r requirements.txt

---

2. Download Pretrained Models

To download the pretrained checkpoints, run:

bash commands/download_pretrained_models.sh

The checkpoints will be saved under ./outputs/.

---

3. Download and Prepare the Dataset

Please follow the official DrivAerNet++ dataset instructions to download the data from Harvard Dataverse.

Specifically:

Drag values

Download Drag Values and place it at:

./data/drivaernet_plus/drag/DrivAerNetPlusPlus_Cd_8k_Updated.csv

3D meshes

Download DrivAerNet++: 3D Meshes, extract all .stl files into:

./data/drivaernet_plus/meshes/3DMeshesSTL

Then preprocess them with:

cd preprocess
bash extract_data.sh

Pressure fields

Download DrivAerNet++: Pressure and extract all .vtk files into:

./data/drivaernet_plus/pressure/PressureVTK

Sketch conditions

Download DrivAerNet++: Sketches / sketches-CLIPasso.zip and extract it into:

./data/drivaernet_plus/condition/matched_sketches-CLIPasso

---

Shape-Physics Variational Autoencoder (SP-VAE)

Training SP-VAE consists of two stages: independent training and joint training.

Stage 1: Independent Training

You may skip this stage if you use the provided pretrained checkpoints in ./outputs/ShapeVAE, ./outputs/PhysDec, and ./outputs/DragDec.

To train your own model, run:

# Shape encoder and decoder
bash commands/train_shapeautoencoder_single_node.sh

# Pressure decoder
bash commands/train_physdec_single_node.sh

# Drag decoder
bash commands/train_dragdec_single_node.sh

---

Stage 2: Joint Training

You may skip this stage if you use the provided pretrained checkpoints in ./outputs/FinetuneAll.

If you want to use your own Stage-1 checkpoints, please set the checkpoint paths in configs/MultiTaskJoint.yaml:

# lines 43-45
shape_model_ckpt: [path to your ShapeVAE checkpoint]
physics_model_ckpt: [path to your PhysDec checkpoint]
drag_model_ckpt: [path to your DragDec checkpoint]

To train your own model, run:

# training
bash commands/train_totalvae_single_node.sh

# testing
# set the checkpoint paths in configs/MultiTaskTest.yaml (lines 43-45)
bash commands/test_totalvae_single_node.sh

---

Physics-Guided Shape Generation

1. Unconditional Generation

Training

You may skip this stage if you use the provided pretrained checkpoints in ./outputs/Diffusion-Unconditional

To train your own model, run:

bash commands/train_uncondiff_single_node.sh

Testing

If you want to test your own checkpoint, set the checkpoint path in configs/Diffusion-uncond-test-givenshape.yaml

# line 186
ckpt_path: [path to your checkpoint]

Then run:

bash commands/run_test_uncond_loop_givenshape.sh

---

2. Conditional Generation

Training

You may skip this stage if you use the provided pretrained checkpoints in ./outputs/Diffusion-Conditional

To train your own model, run:

bash commands/train_condiff_single_node.sh

Testing

bash commands/run_test_condiff_loop_fixdrag.sh cond_diff_CLIPasso last
# if you want to test your own checkpoint, use the following command:
bash commands/run_test_condiff_loop_fixdrag.sh [model_name] [checkpoint_name]

---

Citation

If you find this repository useful, please consider citing:

@article{you2025physgen,
  title   = {PhysGen: Physically Grounded 3D Shape Generation for Industrial Design},
  author  = {You, Yingxuan and Zhao, Chen and Zhang, Hantao and Xu, Mingda and Fua, Pascal},
  journal = {arXiv preprint arXiv:2512.00422},
  year    = {2025}
}