ShapeMorph: 3D Shape Completion via Blockwise Discrete Diffusion

Introduction

This repository is the official pytorch implementation of our paper: "ShapeMorph: 3D Shape Completion via Blockwise Discrete Diffusion".

🚀 Training

❄️ Dataset Preparation

We process our dataset as 3DILG.

⛄ Shape Encoding

torchrun --nproc_per_node=4 run_vqvae.py --output_dir output/vqvae --model vqvae_512 --batch_size 32 --num_workers 40 --lr 1e-3 --disable_eval --point_cloud_size 2048

☃️ Shape Completion

Please revise --vqvae_pth to checkpoint path.

torchrun --nproc_per_node=4 run_diffusion.py --output_dir output/diffusion --model blockwise_diffusion --vqvae vqvae_512 --vqvae_pth output/vqvae_512 --batch_size 16 --num_workers 40 --lr 1e-3 

🛸 Sampling

Please revise --model_pth and --vqvae_pth to the checkpoints path.

python completion.py --model discrete_diffusion --model_pth output/diffsion --vqvae vqvae_512 --vqvae_pth output/vqvae_512 --sample_n 5 

🏁 Results

Multimodal completion on ShapeNet.

Multimodal completion on PartNet.

Multimodal completion on real-scan dataset RedWood.

📬 Contact

Contact Zalfy if you have any further questions.