NV-Generate-MR
Generate synthetic body MRI volumes through NVIDIA’s rflow-mr workflow.
The skill clearly limits use to synthetic MRI, prohibits clinical and production-training use, and discloses output, cache, and Hugging Face/GitHub network effects. It pins the upstream commit, inventories model files, records hashes, and avoids deleting outputs. Deductions reflect Bash access, pip installation and network downloads without dependency hashes or sandbox isolation; ordinary runs lack explicit user confirmation, may write to a default output/cache location, and provide no rollback mechanism.
The wrapper includes argument validation, CUDA and model preflight checks, voxel limits, a cost gate, timeout handling, stderr capture, NIfTI summarization, and structured failure output. Deductions reflect static-only review, reliance on an upstream repository and undisclosed manifest/requirements details, and unconfirmed compatibility among upstream models, configs, and commands. Cross-environment reproducibility and abnormal-path coverage remain limited. Static calibration caps this at 10.
Purpose, exclusions, supported modalities, brain-MRI redirection, input configuration, and output types are documented, and the command interface is usable by Chinese-speaking users. Deductions reflect the NVIDIA GPU requirement, no CPU fallback, and core dependence on Hugging Face/GitHub for setup and downloads. Mainland-China reachability or mirror guidance is absent, and invocation still depends substantially on user judgment about input format and scope.
The skill provides purpose, instructions, prerequisites, limitations, troubleshooting, references, fixtures, and benchmark material; it states the wrapper license and upstream rights constraints and records an upstream commit. Deductions reflect the missing recommended Examples section, only two benchmark tasks, unclear maintenance ownership, absent changelog/version policy, and a possible inconsistency between license metadata and repository-level licensing. Static findings also report unexpected root directories.
The wrapper stages configuration overrides, invokes the upstream inference path, and emits model inventory, configuration, geometry, spacing, affine, intensity, and finiteness checks. The supplied evaluation report shows 85% assisted correctness and 57% assisted effectiveness for Codex. Deductions reflect only two one-shot evaluation tasks and the inability of static review to confirm actual generated outputs; results still require independent quality and legal review, and comparative benefit over direct upstream use is only weakly evidenced. Static calibration caps this at 7.
The source includes a pinned upstream revision, exact commands, structured results, model-file hashes, output hashes, and validation fields; BENCHMARK.md and evals.json provide limited evaluation evidence. Deductions reflect that this review did not execute anything, benchmark coverage is narrow, and no complete CI or committed key-path test evidence is provided. Upstream behavior, model quality, and output correctness cannot be independently reproduced from the supplied files alone. Static calibration caps this at 5.
- Execution uses Bash, installs requirements.txt dependencies, and may download code or models from Hugging Face/GitHub; review network access, dependencies, and model licensing first.
- Outputs are synthetic, image-only MRI volumes and must not be used directly for clinical interpretation, diagnosis, regulatory submission, or production training data.
- The workflow requires an NVIDIA GPU with approximately 16 GB or more of VRAM and has no CPU fallback; mainland-China access and downloads may fail because no mirror or offline path is documented.
- Normal runs create output directories and staged configs without a rollback or cleanup strategy; use an isolated, dedicated run directory.
What does this skill do, and when should you use it?
NV-Generate-MR is a thin NVIDIA medical-AI skill wrapper for generating synthetic body MRI volumes. It delegates inference, sampling, and decoding to the rflow-mr workflow in NVIDIA-Medtech/NV-Generate-CTMR. The skill supports mri, mri_t1, mri_t2, and mri_flair modalities and produces synthetic NIfTI volumes with result metadata and validation checks. It does not produce paired segmentation masks and is not intended for production training data, clinical deployment, or autonomous diagnosis.
It reads a user-supplied JSON model configuration override, stages the upstream configuration, installs the declared runtime requirements, and runs scripts/run_mr.py. When needed, setup can clone a pinned upstream checkout and retrieve model assets from GitHub or Hugging Face; inference runs on an NVIDIA GPU. The wrapper writes generated outputs to the caller-provided directory and emits a synthetic MRI NIfTI volume plus result_json containing configuration, model inventory, geometry, spacing, affine, intensity range, and finite/non-constant data checks.
- A medical-imaging researcher needs synthetic body MRI volumes for exploratory experiments or method development.
- A developer needs to run the pinned rflow-mr inference workflow in an NVIDIA-GPU environment.
- An evaluation workflow needs generated NIfTI geometry, spacing, affine, and data-validity checks recorded as evidence.
- A research team needs synthetic body MRI in the mri_t1, mri_t2, or mri_flair modality.
What are this skill's strengths and limitations?
- Provides the documented upstream rflow-mr entrypoint without requiring custom inference code for normal runs.
- Supports four named MRI modalities: mri, mri_t1, mri_t2, and mri_flair.
- Records useful output geometry, spacing, affine, intensity, and finite-data checks.
- Distributed through the NVIDIA/skills CLI, whose catalog README documents signatures, evaluations, and daily synchronization.
- Generates image-only volumes and does not emit paired segmentation masks.
- Requires an NVIDIA GPU with at least 16 GB VRAM; the upstream path has no CPU fallback.
- The upstream weights are listed as NVIDIA Non-Commercial, so production-training use requires legal and quality review.
- Not intended for clinical deployment, clinical interpretation, autonomous diagnosis, or regulatory submission; the source provides no independent test-suite or performance-benchmark evidence for this skill.
How do you install this skill?
Install the skill through the NVIDIA/skills skills CLI: npx skills add nvidia/skills --skill nv-generate-mr --yes. Installation can target a client with options such as --agent codex or --agent claude-code. The README does not document a separate source-install procedure for this individual skill.
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/scripts/run_mr.py PATH_TO_MR_CONFIG.json --output-dir runs/nv_generate_mr_demo --modality mri_t1 --random-seed 0. Replace PATH_TO_MR_CONFIG.json with the actual request path and use a fresh output directory. The pinned upstream checkout and rflow-mr weights must be available before inference; runtime requires an NVIDIA GPU with at least 16 GB VRAM. Inspect the emitted JSON and verifier guidance after the run.
How does this skill compare with similar options?
For brain MRI synthesis, the upstream README recommends rflow-mr-brain and the separate nv-generate-mr-brain skill; NV-Generate-MR is intended for body MRI rather than that brain-imaging path.