How kaist-visual-ai-group/salad works
Project type and purpose
• It implements a two‐phase part‐level latent diffusion model, with training/inference CLI entry points, a Hydra‐driven config system, core model components, data loaders, and an interactive “SPAGHETTI” UI submodule.
File structure
• Top level
• train.py – CLI entry for training (unconditional/text‐guided).
• setup.py – package installation.
• configs/ – Hydra YAML config hierarchy (model, trainer, paths, logger, callbacks).
• environment.yml – conda spec.
• notebooks/ – Jupyter demos for generation and part completion.
• Core library (salad/)
• data/ – dataset definition.
• model_components/ – building blocks (transformer, LSTM, variance schedule).
• models/ – high‐level model definitions (phase1, phase2, language phases).
• spaghetti/ – interactive UI tool for mesh inference/editing, with own checkpoints, utils, and assets.
• utils/ – shared utilities (logging, mesh I/O, path management, etc.).
README for additional insights
• Confirms two‐stage diffusion (low‐dim extrinsic then high‐dim intrinsic).
• Uses PyTorch + Hydra + CUDA pointnet2 ops.
• Provides training CLI examples and data/checkpoint layout.
• SPAGHETTI is an open‐source dependency embedded under salad/spaghetti/.