Point Cloud Generation

Point cloud generation studies how to synthesize sets of 3D points representing objects, scenes, or sensor returns, supporting reconstruction, design, simulation, and data augmentation. Common approaches include point-cloud diffusion, flow matching, and latent or hierarchical models that separate global shape or layout from local geometry; conditions can include images, text, partial scans, actions, or physical parameters. Evaluation spans shape benchmarks such as ShapeNet and ModelNet, scene datasets such as SemanticKITTI and KITTI-360, and downstream tasks including detection and pose estimation. Current work aims for faster sampling, arbitrary resolution, and control across sensors and large scenes, while balancing geometric fidelity, point coverage, diversity, and reliable downstream gains.