Used for finetuning NV-Generate-CTMR MR-brain diffusion UNet from a NIfTI datalist. Not for clinical or production data approval.
NV-Generate-MR-Brain-Finetune is an NVIDIA MedTech skill for finetuning the NV-Generate-CTMR rflow-mr-brain diffusion UNet from user-supplied NIfTI brain-MRI training volumes. It is explicitly scoped for research and development only and is not for clinical interpretation, regulatory use, or approving synthetic data for production training. It solves the problem of correctly wiring up an upstream, multi-step diffusion training pipeline by acting as a thin, well-documented wrapper rather than a reimplementation.
The wrapper stages the upstream config and environment JSON files locally and delegates execution to the existing upstream scripts for creating training data, training the diffusion model, and optionally running inference; it does not execute the tutorial notebook itself. It exposes the most commonly tuned fields as CLI flags (epochs, modality, GPU count, AMP toggle, model config, run-inference) on a single entrypoint, run_mr_brain_finetune.py. A --preflight mode validates a datalist and upstream script discovery without touching a GPU, and is the recommended first step. Inputs are a MONAI-style datalist and a data base directory; outputs are a finetuned checkpoint, optional inference images, and a result JSON. The skill ships a manifest, output schema validator, tests, fixtures, and a benchmark report from NVIDIA's NVSkills-Eval, which scored the skill 100% on its security dimension.
It targets ML developers and engineers working with medical imaging generative models. Real runs require a CUDA GPU, upstream Python dependencies, and downloaded MR-brain weights; the wrapper may clone the pinned upstream repository from GitHub and fetch model assets from Hugging Face, writing checkpoints and caches under caller-provided and cache directories.
No. The documentation repeatedly states it is for research and development only and must not be used for clinical interpretation, regulatory submission, or approving synthetic data for production training.
Use --preflight, which validates a new datalist and confirms upstream script discovery without using a GPU or launching training. You remove --preflight only when you explicitly want to start real GPU finetuning.
Real training needs an NVIDIA CUDA GPU, the Python packages from the upstream requirements file, and downloaded MR-brain weights. You supply a MONAI-style datalist and a data base directory; the wrapper writes checkpoints, configs, and logs under your output directory.
It expects a local NV_GENERATE_ROOT checkout containing the upstream scripts. If none is supplied it can clone the pinned NV-Generate-CTMR repository from GitHub at a fixed commit and may fetch model assets from Hugging Face, caching under the user's home cache directory.
No. It is deliberately a thin wrapper that stages config glue and delegates to the official upstream scripts for data creation, training, and inference, so behavior matches the upstream tutorial rather than a handwritten reimplementation.
Quick Setup:
.claude/skills/Repository
nvidia/skills