Official implementation of SPLICE: a part-level 3D shape editing framework based on local semantic extraction and global neural mixing.

SPLICE: Part-Level 3D Shape Editing from Local Semantic Extraction to Global Neural Mixing

SPLICE: Part-Level 3D Shape Editing from Local Semantic Extraction to Global Neural Mixing
Jin Zhou, Hongliang Yang, Pengfei Xu, Hui Huang
Pacific Graphics 2025

Abstract

Neural implicit representations of 3D shapes have shown great potential in 3D shape editing due to their ability to model high-level semantics and continuous geometric representations. However, existing methods often suffer from limited editability, lack of part-level control, and unnatural results when modifying or rearranging shape parts.

In this work, we present SPLICE, a novel part-level neural implicit representation of 3D shapes that enables intuitive, structure-aware, and high-fidelity shape editing. By encoding each shape part independently and positioning them using parameterized Gaussian ellipsoids, SPLICE effectively isolates part-specific features while discarding global context that may hinder flexible manipulation. A global attention-based decoder is then employed to integrate parts coherently, further enhanced by an attention-guiding filtering mechanism that prevents information leakage across symmetric or adjacent components.

Key Capabilities

Through this architecture, SPLICE supports a wide range of intuitive part-level editing operations while preserving semantic consistency and structural plausibility:

• Translation, Rotation, & Scaling
• Part Deletion & Duplication
• Cross-Shape Part Mixing

---

🚀 Getting Started

Environment Configuration

Ensure you have Python installed, then run the following commands to set up your environment. This project uses PyTorch 2.6.0 with CUDA 12.6.

# Core PyTorch installation
pip install torch==2.6.0 torchvision==0.21.0 torchaudio==2.6.0 --index-url https://download.pytorch.org/whl/cu126

# Sparse convolution and PyTorch Lightning
pip install spconv-cu120
python -m pip install lightning

# Torch Scatter
pip install torch-scatter -f https://data.pyg.org/whl/torch-2.6.0+cu126.html

# 3D processing, utilities, and logging
pip install vtk==9.3.1
pip install timm einops addict python-dotenv
pip install hydra-core omegaconf trimesh h5py scipy 
pip install pyvista scikit-image libigl wandb pynput

# Flash Attention (requires packaging and ninja first)
pip install packaging ninja
pip install flash-attn --no-build-isolation

Dataset Preparation

1. Clone the Shape-As-Points repository to get the dataset scripts:

git clone https://github.com/autonomousvision/shape_as_points.git
cd shape_as_points
bash scripts/download_shapenet.sh

2. Download the datasets and organize them into your data/ directory:

• Place the downloaded shapenet_psr dataset into data/.
• Download the PartNet dataset and place it into data/.
• Ensure partnet_metadata.json is located directly in the data/ folder (or inside data/shapenet/ depending on your specific code config).

Expected Directory Structure:

data/
├── partnet_metadata.json
├── shapenet/                  # From shapenet_psr
│   ├── {shapenet_id_1}/
│   │   └── psr.npz
│   └── ...
└── {partnet_id_1}/            # PartNet samples
    ├── point_sample/
    │   ├── pts-10000.txt
    │   └── label-10000.txt
    └── ...

🏋️ Training Pipeline

The training process is divided into three sequential stages:

1. Stage 1: Base Model Training Run the main training script. This will generate the initial model checkpoints.

python train.py

Output: Checkpoints will be saved in the checkpoints_splice/ directory. 2. Stage 2: Generate Diffusion Training Data Extract the features/data required to train the diffusion generative model.

python pred_diff.py

Output: Training data is exported to export/prediff/. 3. Stage 3: Diffusion Model Training Train the diffusion model using the data generated in the previous step.

python traindiff.py

Output: Diffusion checkpoints will be saved in checkpoints_splice_diff/.

📥 Pre-trained Models (Optional)

If you want to skip training, you can download pre-trained weights directly from Hugging Face:

huggingface-cli download doudin404/splice --local-dir ./checkpoints --local-dir-use-symlinks False

🧪 Testing & Inference

Once you have trained or downloaded the checkpoints, you can evaluate the models using the following scripts:

• View Reconstruction Results:

python pred.py

• View Random Generation Results:

python samplediff.py

• Interactive Demo: Launch the interactive GUI window. (Note: Operation methods and controls are similar to the SPAGHETTI project).

python demo.py

---

Would you like me to add a section explaining the arguments for the demo.py or train.py scripts, or is this README good to go for your repository?