Release v0.4.0

Features

• Introduced global Point Transformer architecture optimized for Diffusion models.
• Developed a unified Rectified Flow, Laplacian Flow, and Mean Flow training script (train.py) for Point Cloud Generation.
• Implemented memory-efficient Gradient Checkpointing to dramatically reduce VRAM usage during generation training.

Refactors & Build

• Re-architected project structure by migrating all legacy point cloud classification experiments into a dedicated experiments/pc/classification directory.
• Established a new dedicated experiments/pc/generation directory for Flow-based diffusion models.
• Cleaned up legacy top-level point cloud modules (attn.py, block.py, models.py, train_cls.py).
• Updated core dependencies inside pyproject.toml.
• Added .gitattributes configuration and cleared stale Jupyter Notebook outputs.

Examples

• Removed outdated examples for a cleaner core repository.