Tao Generate Od Defects

NVIDIAOfficialCreated by NVIDIA Corporation

Skill
Workflow
Data and AI
Apache-2.0
Medium risk
Has scripts

From the author

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.

— the description by NVIDIA Corporation

AutoGineer summary

Written by AutoGineer from the skill’s own files, not by the author

This skill provides a workflow for generating synthetic object-detection defect images and corresponding COCO-format labels using the AnomalyGenNext 1.1 container. It supports both prepared input plans and native testcase execution. The process includes optional guardrails for content safety and produces validated fine-grained and binary defect annotations. This tool is strictly for synthetic data generation and does not perform model training or integrate outputs directly into training pools.

What it does

Runs AnomalyGenNext 1.1 inference to create synthetic defect images.
Generates fine-grained and binary COCO-format labels for detected defects.
Supports optional guardrails for text screening, image content safety, and face blurring.
Validates input integrity manifests and checkpoint paths before execution.
Produces a validation summary and status report for generated datasets.
Enforces strict container mount requirements for checkpoint and output directories.

When to use it

When you need to generate synthetic defect data for object detection tasks.
When you have prepared input data from the tao-prepare-anomalygennext-inputs workflow.
When you need to create binary or fine-grained COCO labels for anomaly detection.
When you require a controlled generation environment with optional safety guardrails.

Advantages

Automates the complex setup and execution of AnomalyGenNext inference.
Ensures consistent output formatting with validated COCO annotations.
Provides built-in safety mechanisms to screen generated content.
Strict validation of inputs and checkpoints prevents common runtime errors.

Usage examples

“Generate synthetic defects using my prepared input directory.”: The agent executes the generation script using the specified inputs-dir, base-checkpoint, and checkpoint-root, producing synthetic images and COCO labels in the output directory.
“Run AnomalyGenNext generation without guardrails.”: The agent triggers the generation process with the --no-guardrail flag, disabling text screening, image safety, and face blurring, ensuring the requested row count is fully generated.

Install

Download the skill as a ZIP:
https://mp-staging.autogineer.com/agent-skills/nvidia/tao-generate-od-defects/download

Unzip it into the cross-client skills folder of your project:

mkdir -p .agents/skills && unzip tao-generate-od-defects.zip -d .agents/skills

Documentation by NVIDIA Corporation

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.

Files

ATTRIBUTION.mdAdded by AutoGineer939 B
BENCHMARK.md6.9 KiB
LICENSELicence file added by AutoGineer11.2 KiB
SKILL.md3.8 KiB
config/skillspector-baseline.yaml1.4 KiB
evals/evals.json901 B
references/container-runtime.md2.9 KiB
references/execution-contract.md4.5 KiB
references/skill_info.yaml1.8 KiB
scripts/generate_od_defects.pyScript15.0 KiB
scripts/tests/test_generate_od_defects.pyScript10.9 KiB
skill-card.md4.7 KiB
skill.oms.sig5.9 KiB

references/execution-contract.md

AnomalyGenNext OD generation contract

Read this when adapting generation to a new platform or AnomalyGenNext release. This action performs inference only and requires an existing task checkpoint with its matching canonical recipe.

Input modes

Exactly one input mode is allowed:

Prepared mode: --inputs-dir points to a completed tao-prepare-anomalygennext-inputs result containing anomalygen_next_generation_plan.json.
Native mode: pass --input-data-path, --checkpoint, and --recipe; optional dataset and anomaly-type selectors may narrow existing rows.

Every testcase image and mask must exist. Each anomaly_type must be present in the recipe, and the requested count must equal the testcase row count. Prepared-input hashes are verified when the integrity manifest is present. The complete checkpoint root is a separate required input mounted at /workspace/paidf-anomalygen/checkpoints; its fixed Hugging Face and DINOv2 assets are verified before native generation starts.

The wrapper exposes the native Boolean guardrail choice. Guardrails default on without changing the native invocation; disabling them appends --no-guardrail explicitly. --no-guardrail disables the text guardrail, image content-safety path, and face-blur postprocessor together. The validation summary records the selected mode. The pinned framework's default video preset has no active image safety model: its SigLIP-based VideoContentSafetyFilter is disabled because of excessive false positives. Enabling an enforcing image classifier requires a separately validated container preset rather than another argument to this wrapper.

Native stages

From the release source baked into the image, the wrapper invokes generation, optional evaluation and quality refinement when a real-image root is present, and pseudo-labeling. Generation uses torchrun with the requested GPU count; AnomalyGenNext owns distributed row partitioning and rank-zero metadata merge.

Completion accounting

For every dataset:

generated + guardrail_blocked == requested

With --no-guardrail, generated == requested and guardrail_blocked == 0.

Every generated image must have exactly one pseudo-label image record and at least one valid in-bounds annotation. The native COCO retains the exact fine-grained anomaly categories. The binary companion maps every annotation to category id 1 named defect.

The leaf never admits generated images into a detector training pool. That decision belongs to the calling application.

Image-safety reporting

The native guardrail path includes three separate components: text screening, image content-safety screening, and face-blur post-processing. The observed image-safety state applies only to image content-safety enforcement; it does not imply that text screening or face blurring was active. Report component states independently when the runtime exposes them.

Component	Stage	Effect
Text screening	Before generation	A blocklist and text-safety classifier screen the input prompt and any upsampled prompt. Unsafe text can stop the sample before image generation.
Image content-safety screening	After generation	A content-safety classifier evaluates generated visual content and can reject an unsafe result before downstream use.
Face blur	After generation	A face detector locates faces and blurs the detected regions in the output. This is a privacy transformation, not a safe/unsafe verdict, and does not by itself prove image-safety enforcement.

A workflow that records the observed image-safety state must use one of these values:

Value	Meaning
`enforcing`	The runtime positively confirmed that image-safety enforcement was active.
`not_enforcing`	The runtime positively confirmed that image-safety enforcement was inactive.
`unknown`	The wrapper could not determine whether enforcement was active.

These values describe an observed result, not a requested mode. Treat not_enforcing and unknown as unscreened; neither is equivalent to enforcing. A zero guardrail_blocked count does not prove that enforcement ran. This run-level observation is distinct from per-image guardrail results.

Quality metrics

When native evaluation is enabled, interpret FID and nn_score according to the AnomalyGenNext evaluation reference. They are diagnostic signals with sample-count and visual-plausibility limitations, not standalone acceptance evidence.

Requirements and compatibility

Works with: Claude Code, Claude apps, OpenAI Codex, GitHub Copilot, Cursor, Gemini CLI, Kiro, OpenCode, Goose, Amp, Roo Code, JetBrains Junie, Factory Droid, Devin (ex-Windsurf), Mistral Vibe, OpenHands, Letta

“Works with” means the tool reads the open Agent Skills format; a skill written for one tool may use features others lack.

Needs: Declares allowed tools, Needs Python

The author says: Requires the AnomalyGenNext 1.1 container, CUDA GPUs, base and task checkpoints.

Security

What our scanner noticed· Medium risk

Large base64-encoded blobs
Scripts that your agent may run

This skill includes scripts that your agent may run. Read them before you use it.

A static scan is not a guarantee. Skills are third-party content: review a skill before you let an agent use it.

Attribution and licence

Created by

NVIDIA Corporation

Publisher

NVIDIA (Official)

Source

NVIDIA/skills · skills/tao-generate-od-defects · commit fce3cca

Licence

Apache License 2.0 (redistribution allowed)

LICENSE (added by AutoGineer from LICENSE-APACHE)
ATTRIBUTION.md (written by AutoGineer)

Synced

2026-11-08

AutoGineer is not affiliated with or endorsed by NVIDIA Corporation or NVIDIA.

Download ZIP

The ZIP holds the skill’s files, its licence files and an ATTRIBUTION.md. Download tao-generate-od-defects.zip