SuperPC: A Single Diffusion Model for Unified Point Cloud Processing

[Paper] [Code] [Homepage]

The official code repository for the CVPR 2025 paper "SuperPC: A Single Diffusion Model for Unified Point Cloud Processing".

News

1. Added dataset preparation code and instructions
2. Uploaded the pretrained weights
3. Added evaluation and single-sample testing code

Future Plans

• Add detailed training pipeline
• Possibly add prepared datasets if permission is granted by the original dataset providers
• Add the flow-matching based training pipeline and pretrained weights

Installation

Run the setup script:

./setup_env.sh

Datasets Preparation

Follow the detailed dataset preparation guide to download the three raw datasets and prepare ground-truth point clouds and images.

1. ShapeNet dataset
2. TartanAir dataset - separate sub-dataset list
3. KITTI-360 dataset

Model Zoo

Download all pretrained weights from: https://drive.google.com/drive/folders/1FrQtm8LBVrbdRT4Xs87rIZpJ9nYaTqcG?usp=drive_link

Evaluation and Testing (Combined Tasks)

1. ShapeNet Dataset

Use test_superpc.py for single-sample or split-based testing with optional output saving.

[Single sample] (assuming you have already prepared the ShapeNet dataset):

airplane: /02691156/10aa040f470500c6a66ef8df4909ded9

python test_superpc.py \
  --dataset shapenet \
  --cat-id 02691156 \
  --model-id 10155655850468db78d106ce0a280f87 \
  --mode easy \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --vision_image_dir /path/to/shapenet/render/images \
  --num_points 2048 \
  --target_num_points 8192 \
  --use_input_scout_fill true \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

[Whole ShapeNet test split] (omit --cat-id and --model-id):

python test_superpc.py \
  --dataset shapenet \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --vision_image_dir /path/to/shapenet/render/images \
  --num_points 2048 \
  --target_num_points 8192 \
  --use_input_scout_fill true \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

2. TartanAir Dataset

Use test_superpc.py for single-sample or split-based testing with optional output saving.

[Single sample] (assuming you have already prepared the TartanAir dataset):

Single TartanAir sample by metadata path:

python test_superpc.py \
  --dataset tartanair \
  --tartanair_root /data_sair/tartanair_maps \
  --metadata_path /path/to/tartanair_maps/hospital/Easy/P000/000000/metadata.json \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --num_points 11520 \
  --target_num_points 46080 \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

TartanAir split testing (--eval_split: train, val, test, or all):

python test_superpc.py \
  --dataset tartanair \
  --tartanair_root /data_sair/tartanair_maps \
  --eval_split test \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --num_points 11520 \
  --target_num_points 46080 \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

3. KITTI-360 Dataset

Use test_superpc.py for single-sample or split-based testing with optional output saving.

[Single sample] (assuming you have already prepared the KITTI-360 dataset):

Single KITTI-360 sample by metadata path:

python test_superpc.py \
  --dataset kitti360 \
  --kitti360_root /path/to/kitti360_maps/submaps \
  --metadata_path /path/to/kitti360_maps/submaps/2013_05_28_drive_0000_sync/image_00/0000000090/metadata.json \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --num_points 11520 \
  --target_num_points 46080 \
  --use_input_scout_fill true \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

KITTI-360 split testing (--eval_split=test maps to the val split because KITTI-360 currently uses train/val partitioning):

python test_superpc.py \
  --dataset kitti360 \
  --kitti360_root /path/to/kitti360_maps/submaps \
  --eval_split test \
  --ckpt_path /path/to/your_checkpoint.pth \
  --use_vision_conditioning true \
  --num_points 11520 \
  --target_num_points 46080 \
  --seed 21 \
  --sampling_steps 25 \
  --save_pc true

Acknowledgement

The codebase is built upon: MinkowskiEngine, PUFM, DiffusionPC, and DCD. We appreciate the authors' excellent work.

Reference

@inproceedings{du2025superpc,
  title={SuperPC: a single diffusion model for point cloud completion, upsampling, denoising, and colorization},
  author={Du, Yi and Zhao, Zhipeng and Su, Shaoshu and Golluri, Sharath and Zheng, Haoze and Yao, Runmao and Wang, Chen},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={16953--16964},
  year={2025}
}