Synthetic Brain MRI Generator
Generate synthetic brain MRI volumes through NVIDIA’s documented workflow.
The skill discloses writes to the output directory, model caching, dependency installation, and GitHub/Hugging Face access, while restricting use to synthetic images and prohibiting clinical use. It includes a cost-confirmation gate, output validation, and a recommended pinned upstream commit. Points are deducted because Bash, pip installation, and network downloads are broad capabilities; fine-grained isolation, sensitive-data handling, and complete rollback are not specified. The script also writes to the upstream directory when --output-dir is omitted, which conflicts somewhat with the documented caller-provided directory model.
The documentation, wrapper, and documented command are mostly consistent for the main path. The script provides argument validation, CUDA and memory preflight checks, model inventory, timeout handling, stderr capture, and NIfTI output checks. Points are deducted because GPU, upstream-code, package, and model availability are external prerequisites, failure coverage is limited, and behavior can drift with upstream dependencies. This was a static review, so the score remains below the static ceiling of 10.
The target scenario, configuration input, output directory, modality list, and major non-fit boundaries are fairly clear, including skull-stripped FLAIR. Points are deducted because the workflow requires an NVIDIA GPU with roughly 16 GB or more VRAM, depends on GitHub/Hugging Face downloads, and does not document Chinese-language interaction support. Core setup may depend on overseas services that can be difficult to reach reliably from mainland China, and invocation boundaries remain only partly specified.
The skill includes Purpose, Instructions, Prerequisites, Limitations, Troubleshooting, Usage, and reference material. Naming and configuration keys are fairly stable, and author plus Apache-2.0 information is present; the repository also supplies license, security-reporting, and benchmark material. Points are deducted for the missing recommended Examples section, unclear skill-level versioning, changelog, maintenance ownership, and update path, as well as the supplied NOASSERTION license metadata and benchmark schema findings.
The wrapper covers the core synthetic brain-MRI generation flow and adds modality selection, deterministic seeds, configuration overrides, cost estimates, and output geometry/finite-data checks. The benchmark file reports two positive tasks with some correctness and discoverability gains. Points are deducted because the output is image-only, segmentation masks are not produced, GPU and runtime costs are substantial, and static review cannot confirm directly usable final NIfTI results; the score therefore stays within the static ceiling of 7.
The files provide a pinned upstream revision, documented command shape, configuration fixtures, output-summary fields, and benchmark results, giving some auditability. Points are deducted because the benchmark has only two positive tasks, no negative tasks, and no independently reproducible full test evidence or cross-source corroboration. No code was executed in this review, so the score remains below the static ceiling of 5.
- Do not use the generated MRI for clinical interpretation, diagnosis, regulatory submission, or production training data without independent quality review.
- Before running, verify the input configuration, output directory, model weights, Python dependencies, and NVIDIA GPU; first-time setup may require GitHub and Hugging Face access.
- Use an explicit fresh output directory and retain the JSON, stderr, and validation results; do not rely on the default output location.
What does this skill do, and when should you use it?
This skill generates synthetic brain MRI volumes with NVIDIA-Medtech/NV-Generate-CTMR’s rflow-mr-brain workflow. It uses the repository’s Python wrapper for configuration staging, output management, and validation rather than reimplementing diffusion sampling or autoencoder decoding. Execution requires an NVIDIA GPU with at least 16 GB of VRAM and has no CPU fallback. Outputs are image-only synthetic volumes without paired segmentation masks and are not intended for production training data, clinical deployment, or diagnosis.
Reads a user-supplied model configuration override, installs the upstream requirements, and invokes scripts/run_mr_brain.py, which delegates inference to the upstream scripts.diff_model_infer workflow. It supports documented MRI modalities, writes generated NIfTI volumes under the caller-provided --output-dir, and emits result_json containing the staged configuration, model inventory, upstream command, output geometry, spacing, affine, intensity range, and finite-data/non-constant checks. It may cache model assets under ~/.cache/huggingface/ and contact huggingface.co or github.com during setup.
- Medical-imaging researchers need synthetic brain MRI volumes for exploratory work or research prototypes.
- MONAI or NVIDIA medical-AI users need to run a documented generator for mri_t1, mri_t2, mri_flair, or another supported modality.
- Engineers need configuration, model, geometry, and validation evidence recorded for reproducible experiments.
- Users need repeatable sampling and can set --random-seed while keeping outputs in a fresh run directory.
What are this skill's strengths and limitations?
- Uses a documented wrapper, so normal runs do not require handwritten inference code.
- Supports multiple brain MRI modalities, including mri_t1, mri_t2, mri_flair, mri_swi, and skull-stripped variants.
- Records configuration, model, output geometry, and data-validity checks for experiment review.
- Supports traceability through a pinned upstream commit and an explicit random seed.
- Requires an NVIDIA GPU with at least 16 GB of VRAM and has no CPU fallback.
- Produces image-only synthetic volumes and no paired segmentation masks.
- Generated volumes are not suitable as production medical-model training data without independent quality review.
- Not for clinical interpretation, clinical deployment, autonomous diagnosis, or regulatory submission.
- Setup may require network access for code, dependencies, and model downloads and may create cache and output-directory side effects.
How do you install this skill?
Install the specific skill with the NVIDIA skills CLI: npx skills add nvidia/skills --skill nv-generate-mr-brain --yes. The source does not state a price. Before running, provide an NVIDIA GPU, the Python dependencies, and the upstream model weights. If NV_GENERATE_ROOT is not already available, the documented setup creates an upstream checkout pinned to commit 61c4ec709b84cad468852243c48e250bec732074 and downloads the rflow-mr-brain weights.
How do you use this skill?
From the repository root, run: export NV_GENERATE_ROOT="${NV_GENERATE_ROOT:-$HOME/.cache/nvidia-skills/upstreams/NV-Generate-CTMR-61c4ec7}" && python -m pip install -r "$NV_GENERATE_ROOT/requirements.txt" && python skills/nv-generate-mr-brain/scripts/run_mr_brain.py PATH_TO_MR_BRAIN_CONFIG.json --output-dir runs/nv_generate_mr_brain_demo --modality mri_t1 --random-seed 1234. Replace PATH_TO_MR_BRAIN_CONFIG.json with the actual configuration path; do not delete or pre-create the output directory. Inspect the wrapper JSON and paired verifier guidance before treating a run as evidence.