Official repository for ControlFlow3D: Distilling Multi-View Knowledge into Latent Flow Matching for Point Cloud Upsampling.
Code coming soon. The training and inference code will be released in this repository. Stay tuned!
Download ShapeNetPU and read the dataset documentation on Hugging Face
ShapeNetPU is a ShapeNet-derived dataset for point cloud upsampling, covering 13 object categories.
- 35,762 shapes: 30,395 for training and 5,367 for testing.
- Paired point clouds: 2,048-point sparse inputs and 8,192-point dense ground truth.
- Multi-view training data: eight rendered views per training shape, with view lists and rendering metadata.
- Evaluation geometry: reference meshes for the test split.
The dataset card provides the directory structure, category-level splits, download instructions, file-format details, and loading examples.
ShapeNetPU is derived from ShapeNet. Please follow the dataset terms of use and the applicable upstream licenses.