We have packaged our core functionalities into a library called libpsc, which you can install quite easily:

pip install trimesh # dependencies
pip install libpsc

Supported platforms:

• python 3.8 ~ 3.13
• linux, windows

You can also find all the wheels here. The source code is also available.

We also provide a "progressive simplicial complex" viewer to visualize the simplification and generation process. We have provided two pre-built executables that should run perfectly on Windows and Linux, which can be downloaded from here.

You first need to install some dependencies

pip install fire trimesh easygui libpsc # dependencies

# clone the repo that contains the executable
git clone https://github.com/Karbo123/armesh.git --depth=1
cd armesh

For example, to use the executable for visualizing assets/example-chair.psc, you may need to:

python vis/vis.py assets/example-chair.psc

You may need to properly set up the code environment for running experiments.

We highly suggest using Linux for training experiments because some used packages have poor support for Windows systems. To set up the packages, you may need the following dependencies, which can be installed through pip. The package versions listed below are just suggestions; other versions might work as well.

torch               2.6.0+cu126
pytorch-lightning   2.5.1
flash_attn          2.7.4.post1
deepspeed           0.16.6
wandb               0.19.9
trimesh             4.6.6
numpy               2.2.4
tqdm                4.67.1
fire                0.7.0
joblib              1.4.2
hydra-core          1.3.2
omegaconf           2.3.0
torchmetrics        1.7.1
retrying            1.3.4
numba               0.61.2
datrie              0.8.2

We provide 100 mesh examples in this folder. We will now present a minimalist experiment demonstrating how we can learn from this tiny dataset.

You can follow the steps below for the experiments:

# simplify the mesh, reverse the sequence
# write psc files to disk
# write "object_info.json"
python exp/1/1-preprocess-dataset.py

# tokenize each mesh, and write "op_tok.json"
python exp/1/2-tokenize.py

# learn vocabulary for tokenization compression ("vocab-bit-high.json" and "vocab-code-vfe.json")
python exp/1/3-learn-vocab.py --name="bit-high" --num_new_vocab=4096  --id_new_start=523  --range_start=2 --range_end=5 
python exp/1/3-learn-vocab.py --name="code-vfe" --num_new_vocab=11763 --id_new_start=4619 --range_start=8 --range_end=-1

# merge vocabulary and compressively tokenize each mesh ("vocab.json" and "op_tok_bpe.jsonl")
python exp/1/4-vocab-compress.py

# start training (some configurations are in "exp/1/cfg.yaml")
python exp/1/5-train.py

# unconditional generation (some configurations are in "exp/1/cfg.yaml")
# please change the path to checkpoint
python exp/1/6-gen.py resume=out/exp-1--YYYY-MM-DD--HH-MM-SS/ARMesh/abcdefgh/checkpoints/best.ckpt

If you find our work useful or use any part of our code, please consider citing our paper; your recognition means a great deal to us!

@misc{lei2025armeshautoregressivemeshgeneration,
      title={ARMesh: Autoregressive Mesh Generation via Next-Level-of-Detail Prediction}, 
      author={Jiabao Lei and Kewei Shi and Zhihao Liang and Kui Jia},
      year={2025},
      eprint={2509.20824},
      archivePrefix={arXiv},
      primaryClass={cs.GR},
      url={https://arxiv.org/abs/2509.20824}, 
}