How lumalabs/imm works
Project type and purpose
• This is a research‐oriented generative model codebase implementing “Inductive Moment Matching.”
• It provides both a training pipeline for diffusion‐like models (under training/) and an inference/sampling script (generate_images.py).
• It relies on Hydra‐style configs (configs/*.yaml), utility libraries (dnnlib, torch_utils), and PyTorch.
File structure
• Top‐level:
• generate_images.py – entry point for sampling
• env.yml – conda environment spec
• configs/ – Hydra config files for different datasets
• dnnlib/ – general utilities (logging, config parsing)
• torch_utils/ – PyTorch utilities (distributed training, persistence, stats)
• training/ – model definitions (UNet, transformers, encoders, preconditions)
• assets/, model_card.md, README.md, license files
• No web front end or external microservices—it’s a self‐contained ML codebase.
• Key external dependency: Hugging Face for checkpoint storage.
README for additional insights
• Shows sampling command: python generate_images.py --config-name=... eval.resume=...
• Describes two modes of time embedding (“s” vs. “t-s”) for ImageNet configs
• Checklist hints that training scripts are still pending (but training modules exist)