nv-generate-mr-brain

The nv-generate-mr-brain skill is a thin wrapper for generating synthetic, image-only brain MRI volumes using NVIDIA's NV-Generate-CTMR rflow-mr-brain model. It solves the problem of correctly invoking a complex upstream diffusion/autoencoder inference pipeline without reimplementing it: the wrapper owns config staging, output paths, and validation, and delegates all sampling and decoding to the pinned upstream NVIDIA-Medtech/NV-Generate-CTMR repository. It is explicitly for research and engineering-verification workflows and not for production training data.

The skill exposes a single entrypoint, scripts/run_mr_brain.py, driven by a model config JSON with an output directory, modality (for example mri_t1), optional random seed, and an optional confirmation flag. It documents a manifest (skill_manifest.yaml) that declares inputs and outputs, prescribes running the documented command exactly (including a pip install of the upstream requirements so fresh environments do not fail on missing packages like nibabel or MONAI), and forbids adding rm, mkdir, or cleanup of the output directory since the wrapper manages it. Side effects include writing NIfTI volumes and a result JSON under the caller's output directory, optionally caching model assets under the HuggingFace cache, and contacting huggingface.co or github.com during setup to fetch model weights. A reference file covers field-of-view settings and the model-only weight download.

Target users are developers and researchers who need synthetic brain MRI volumes for R&D and verification. The skill is clear about limitations: it produces image-only volumes with no segmentation masks, outputs are not safe as production training data without independent quality review, and it is not for clinical deployment, interpretation, autonomous diagnosis, or regulatory submission.