nv-generate-vae-finetune

This NVIDIA MedTech skill wraps the finetuning of the NV-Generate-CTMR MAISI VAE (variational autoencoder) from user-supplied CT or MRI NIfTI training volumes. It solves the problem of running an upstream training workflow that is otherwise only documented in a tutorial notebook: rather than executing the notebook, the wrapper stages the required config and datalist glue locally and calls upstream helper APIs, exposing the most-tuned training fields as CLI flags. Inputs are a MONAI-style datalist and a data base directory; outputs are autoencoder and discriminator checkpoints plus a result JSON.

The primary entrypoint, `run_vae_finetune.py`, accepts a datalist, data base directory, output directory, and flags for epochs, modality, patch size, and a `--preflight` mode. The skill strongly recommends running `--preflight` first to validate a new datalist without a GPU or training, removing it only when the user explicitly wants to launch GPU finetuning. Prerequisites include an explicit `NV_GENERATE_ROOT` pointing at a local upstream checkout (or an automatic search of a workbench cache), an optional `CUDA_VISIBLE_DEVICES`, a CUDA GPU for real training, upstream Python requirements plus `lpips`, and downloaded VAE weights unless training from scratch. Documented side effects are writing staged configs, checkpoints, TensorBoard logs, and summaries under the output directory, caching under the upstream checkout and the Hugging Face/torch caches, and contacting huggingface.co, github.com, and download.pytorch.org during setup. The included skill card marks it research-and-development only with no required credentials.

Target users are developers and medical-imaging engineers finetuning the MAISI VAE for image-synthesis research. The skill states it is not for clinical interpretation, regulatory use, or approving synthetic data for production training.