← LeaderboardThe Skill Leaderboard

tao-generate-od-defects

Run AnomalyGenNext 1.1 inference from a native testcase or a prepared generation plan, then publish validated fine-grained and binary COCO labels. Use for synthetic object-detection defect generation, not model training.

Published by NVIDIA ↗NVIDIA/skills

Try your first task

tao-generate-od-defects

Start a fresh conversation. Replace the highlighted placeholders.

Use tao-generate-od-defects to generate one bounded defect batch from [completed prepared-input directory or native testcase/checkpoint/matching recipe], with [Cosmos3-Nano DCP base], [complete checkpoint root], [GPU allocation] and [new temporary/persistent output paths]. Read execution/runtime contracts and use tao-launch-workflow to review exact image, mounts, runtime and GPU shape before submission. Keep guardrails enabled; require all corresponding tokenizer, processor, DINOv2 and safety assets at the canonical checkpoint mount, then generate offline. Verify prepared hashes, testcase image/mask/type/count agreement, generated plus guardrail-blocked counts, and valid positive in-image COCO boxes/declared categories. Return native fine-grained and binary-defect labels, validation/status artifacts and training_pool_mutated=false; do not admit outputs into detector training.

Required inputs

  • Verified testcase/prepared inputs, requested row count and output paths

User-approved starter prompt based on a pinned source version; check current skill requirements before running. Task evidence ↗

SKILL.md

name: tao-generate-od-defectsdescription: >-  Run AnomalyGenNext 1.1 inference from a native testcase or a prepared  generation plan, then publish validated fine-grained and binary COCO labels.  Use for synthetic object-detection defect generation, not model training.license: Apache-2.0compatibility: Requires the AnomalyGenNext 1.1 container, CUDA GPUs, base and task checkpoints.metadata:   author: NVIDIA Corporation  version: "0.1.0"allowed-tools: Read Bashtags: [tao, data, anomalygen-next, object-detection, synthetic-data, coco]

Generate AnomalyGenNext OD Defects

This leaf uses the upstream generator and pseudo-labeler in the pinned public container. It does not prepare detector gaps, run AMP, train AnomalyGenNext, or append outputs to a detector training set. Read references/execution-contract.md when adapting input or completion behavior, and references/container-runtime.md before submission.

Inputs

Pass either:

  • --inputs-dir pointing to the completed output of tao-prepare-anomalygennext-inputs; or
  • --input-data-path, --checkpoint, and matching --recipe for a native AnomalyGenNext testcase.

Also provide the Cosmos3-Nano base checkpoint, the complete AnomalyGenNext checkpoint root, and one or more GPUs. Every testcase image and aligned mask must exist, every anomaly type must occur in the recipe, and the requested row count must match the testcase. When a prepared integrity manifest exists, every recorded hash is verified before generation.

Run

Invoke tao-launch-workflow, review the exact image, mounts, GPU shape, runtime, and output directory, then submit the generate action:

scripts/generate_od_defects.py \
  --inputs-dir /results/prepared_inputs \
  --base-checkpoint /models/Cosmos3-Nano \
  --checkpoint-root /workspace/paidf-anomalygen/checkpoints \
  --output-dir /temporary/generation \
  --published-root /persistent/generation \
  --no-guardrail \
  --num-gpus 1

When execution uses temporary storage, --published-root records the persistent locations that will contain the saved results.

Guardrails are enabled by default and preserve the native invocation without a guardrail flag. The wrapper appends --no-guardrail only when explicitly disabled. --no-guardrail disables text screening, image content-safety screening, and face blurring together. The checkpoint root is mounted over the image's canonical /workspace/paidf-anomalygen/checkpoints tree and validated before GPU work. The Nano tokenizer, Edge processor, and DINOv2 assets are always required; guardrail-enabled runs additionally require the Qwen3Guard and Cosmos Guardrail repositories. Generation is forced offline after validation. The selected platform must honor the declared checkpoint_root.container_path; the generation leaf rejects a different mount point. The base-checkpoint argument remains separate: it is the parent containing checkpoint.json and the model/ checkpoint directory.

Completion

For each dataset, generated + guardrail_blocked must equal requested rows. With --no-guardrail, generated must equal requested rows and guardrail_blocked remains zero. The validation summary records the selected guardrail mode. The pseudo-label count must match generated images, each image needs an annotation, categories must be declared anomaly types, and every COCO bbox must be positive and inside its image. The action emits:

DATASET/raw/
DATASET/pseudo_labels/coco_annotations.json
pseudo_labels/coco_annotations.json
pseudo_labels/coco_annotations_od_defect.json
validation_summary.json
status.json

The native COCO preserves TEXTURE+DEFECT categories. The binary file maps all annotations to defect. Only a calling application may admit these outputs to training; this leaf always reports training_pool_mutated=false.