PAIDF AnomalyGen
Fine-tune, generate, evaluate, and refine synthetic anomaly images.
The document limits the declared tools to Read and Bash and provides a non-root container, mount-permission preflight, path validation, and confirmation before writing the training config. However, it performs Docker and GPU operations, downloads about 140 GB of checkpoints, and passes HF_TOKEN without sufficient disclosure of token exposure, network data flow, cost confirmation, rollback, or cleanup, so points are deducted.
The phases, parameter validation, output verification, and several failure modes are documented in useful detail. Static evidence provides no tests or reproducible CI covering this skill's key paths, while background training, long-running execution, and recovery from some failures still shift work to the user, so the score remains below the static ceiling.
The audience, three modes, inputs, outputs, and several non-fit conditions are reasonably clear. The skill nevertheless depends tightly on a CUDA GPU, Docker/NVIDIA Container Toolkit, a specific container, and external data services, with no Chinese-language or mainland-China reachability guidance; core downloads may be inaccessible, so points are deducted.
The documentation includes a phase table, quick start, parameter tables, reference-document routing, error handling, output layout, Apache-2.0 licensing, and version 0.1.0. It lacks standard Instructions/Examples sections, an author email, a changelog, and a clearly stated maintenance owner/update path; the supplied benchmark also reports hierarchy and schema findings, so points are deducted.
The skill describes an end-to-end workflow covering fine-tuning, generation, evaluation, search, filtering, regeneration, and declared CSV/log outputs. However, there are no statically verifiable representative outputs for the core pipeline, the workflow is costly, and results still require interpretation; the benchmark covers only a planning task rather than actual pipeline completion, so the score is limited.
A committed benchmark report and evaluation specification provide some evidence, with limited reported security, correctness, and discoverability results. There is only one non-executing planning task and no tests, CI, or independent corroboration covering PAIDF's key phases, preventing a higher static verifiability score.
- Before running, explicitly confirm GPU, disk, time, and roughly 140 GB of download costs, and avoid exposing HF_TOKEN through shell history or container inspection in untrusted environments.
- Core preparation and checkpoint retrieval depend on Hugging Face, Roboflow, GitHub, and a container image; mainland-China network restrictions may prevent completion.
- The document launches training in the background but does not provide a complete process-monitoring, cancellation, checkpoint-recovery, or resource-cleanup procedure.
- The benchmark contains only one planning task and does not establish that the actual generation, evaluation, search, and regeneration pipeline runs successfully.
What does this skill do, and when should you use it?
This skill provides the full PAIDF AnomalyGen multi-phase pipeline for fine-tuning on a new anomaly dataset, generating synthetic anomaly images, and evaluating output quality. It supports full, finetune_only, and inference_only modes. The pipeline can search per-sample guidance and crop_ratio parameters, then filter and regenerate samples by nn_score. It requires Docker, NVIDIA Container Toolkit, a CUDA GPU, and may download about 140 GB of pretrained checkpoints.
The skill reads a dataset root, defect-specification JSONL, and optionally an existing checkpoint; verifies or downloads checkpoints; prepares training and inference JSONL; and runs fine-tuning, SDG generation, evaluation, per-sample parameter search, assembly, filtering, and regeneration. It executes shell scripts and Python utilities in Docker or the supported host environment. Outputs include original, rounds, searched, and regens directories, with files such as SDG_result.csv, per_sample.csv, eval.log, search_summary.csv, and regen_summary.csv.
- A computer-vision team fine-tunes AnomalyGen on a new dataset containing texture, text, or CAD defects.
- A data engineer reuses an existing fine-tuned checkpoint to generate anomaly images without running Phase 1.
- A quality team evaluates generated images using nn_score, mnn_score, and FID outputs.
- A researcher searches guidance and crop_ratio values separately for each sample.
- An operator filters low-scoring samples by nn_threshold and regenerates them automatically.
What are this skill's strengths and limitations?
- Covers fine-tuning, generation, evaluation, search, assembly, filtering, and regeneration in one workflow.
- Supports checkpoint reuse through inference_only as well as fine-tuning-only execution.
- Produces per-sample and aggregate evaluation artifacts for verification.
- Supports 2b and 14b models, plus multi-GPU fine-tuning and SDG.
- Requires a CUDA GPU, Docker, and NVIDIA Container Toolkit; host execution additionally requires the cosmos-predict2 environment.
- Checkpoint downloads can total about 140 GB and require HF_TOKEN.
- Execution depends on dataset structure, defect specifications, mask files, and mount permissions.
- The source provides no test-suite or broader platform-coverage evidence for this individual skill.
How do you install this skill?
Install the skill through the repository README's skills CLI flow:npx skills add nvidia/skills --skill paidf-anomalygen --yes
The skill becomes available when the agent next loads skills and encounters a relevant task. The source does not require separately cloning the skill directory.
How do you use this skill?
Ask the agent to “fine-tune AnomalyGen,” “generate anomaly images,” “run PAIDF SDG,” “evaluate SDG output quality,” or “run per-sample search.” Before execution, provide mode, name, dataset_dir, defect_spec, and num_SDG; inference_only additionally requires both checkpoint_dir and step. Run commands from the repository root with ANOMALYGEN_SCRIPTS exported. Full mode requires reading both the fine-tuning and inference references, and the pipeline is intended to run without mid-run pauses.