Scaling Artist Mesh Generation via Local-to-Global Assembly

CVPR 2026

Rui Xu¹, Tianyang Xue¹, Qiujie Dong¹, Le Wan², Zhe Zhu², Peng Li³, Zhiyang Dou¹, Cheng Lin⁴, Shiqing Xin⁵, Yuan Liu³, Wenping Wang⁶, Taku Komura¹

Affiliations: ¹ The University of Hong Kong ² Tencent Visvise
³ Hong Kong University of Science and Technology ⁴ Macau University of Science and Technology ⁵ Shandong University ⁶ Texas A&M University

🚀 Official code of MeshMosaic: Scaling Artist Mesh Generation via Local-to-Global Assembly

              

🚀 We are preparing the codebase for public release. Stay tuned!

📋 Release Todo List

• Release pretrained checkpoints
• Release inference code
• Release data preprocessing code
• Release training code

Environment Setup

We tested on A100, A800 and H20 GPUs with CUDA 12.4 and CUDA 11.8. Follow the steps below to set up the environment.

1) Create Conda env and install PyTorch (CUDA 12.4)

conda create -n MeshMosaic python=3.12 -y
conda activate MeshMosaic

# PyTorch 2.5.1 + CUDA 12.4
pip install --index-url https://download.pytorch.org/whl/cu124 \
  torch==2.5.1+cu124 torchvision==0.20.1+cu124 torchaudio==2.5.1+cu124

# Core dependencies
pip install -U xformers==0.0.28.post3
pip install torch-cluster -f https://data.pyg.org/whl/torch-2.5.1+cu124.html
pip install packaging

2) Install FlashAttention and custom kernels

git clone https://github.com/Dao-AILab/flash-attention
cd flash-attention
python setup.py install

# If you encounter build issues, try FlashAttention==2.8.0 or 2.7.3

cd csrc/rotary && pip install .
cd ../layer_norm && pip install .
cd ../xentropy && pip install .

3) Install remaining Python packages

pip install pymeshlab jaxtyping boto3 trimesh beartype lightning safetensors \
  open3d omegaconf sageattention triton scikit-image transformers gpustat \
  wandb pudb
pip install libigl

Note: For CUDA 11.8 users, install the corresponding PyTorch/cu118 wheels and compatible torch-cluster build.

Usage

• The script sample.sh demonstrates mesh generation. Input is an OBJ where each part is stored as a distinct connected component within a single file (see folder input_pf).
• Pre-trained weights are available on Hugging Face: here.

Or run directly with the following command:

torchrun --nproc-per-node=1 --master_port=61107 sampleGPCBD.py \
  --model_path "ckpt/final.bin" \
  --steps 40000 \
  --input_path input_pf \
  --output_path output \
  --repeat_num 4 \
  --uid_list "" \
  --temperature 0.5

If you do not have connected-component inputs, you can use the code in the PartField folder to convert your own mesh into an OBJ with semantic segmentation. Below is a concise workflow.

PartField environment setup

• Option A: Manual installation

conda create -n partfield python=3.10 -y
conda activate partfield

conda install -y nvidia/label/cuda-12.4.0::cuda

pip install psutil
pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 \
  --index-url https://download.pytorch.org/whl/cu124
pip install lightning==2.2 h5py yacs trimesh scikit-image loguru boto3
pip install mesh2sdf tetgen pymeshlab plyfile einops libigl polyscope \
  potpourri3d simple_parsing arrgh open3d
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.4.0+cu124.html

sudo apt-get update && sudo apt-get install -y libx11-6 libgl1 libxrender1
pip install vtk

• Option B: Use provided environment file

conda env create -f environment.yml
conda activate partfield

Prepare checkpoint directory

mkdir -p model

Download the pretrained checkpoint: Trained on Objaverse.

Note: Due to licensing restrictions, the model also trained on PartNet cannot be released.

Extract Feature Field

Run from the PartField project directory.

python partfield_inference.py -c configs/final/demo.yaml \
  --opts continue_ckpt model/model_objaverse.ckpt \
         result_name partfield_features/objaverse \
         dataset.data_path data/objaverse_samples

Part segmentation

We use agglomerative clustering for mesh part segmentation.

python run_part_clustering.py \
  --root exp_results/partfield_features/objaverse \
  --dump_dir exp_results/clustering/objaverse \
  --source_dir data/objaverse_samples \
  --use_agglo True \
  --max_num_clusters 30 \
  --option 0

Split connected components

python split_connected_component.py

Please follow our progress for updates. Training code and more resources will be released soon!

Ack

Our code is based on these wonderful works:

• DeepMesh
• PartField
• BPT
• Michelangelo

📚 Citation

If you find this work useful, please cite our paper:

@article{xu2025meshmosaic,
  title={MeshMosaic: Scaling Artist Mesh Generation via Local-to-Global Assembly},
  author={Xu, Rui and Xue, Tianyang and Dong, Qiujie and Wan, Le and Zhu, Zhe and Li, Peng and Dou, Zhiyang and Lin, Cheng and Xin, Shiqing and Liu, Yuan and others},
  journal={arXiv preprint arXiv:2509.19995},
  year={2025}
}