physical-ai-defect-image-generation

Physical AI Defect Image Generation orchestrates end-to-end defect image generation, augmentation, and labeling pipelines for Automated Optical Inspection (AOI) datasets using NVIDIA Cosmos AnomalyGen (a Cosmos-Predict2-2B model fine-tuned per use case) running on OSMO. It addresses the cold-start data problem in visual inspection: teams that lack real defect imagery for PCBA, metal surface, or glass inspection can synthesize labeled defect datasets from CAD/USD scenes and image-edit augmentation, then transition to inference and labeling on real photos.

The skill governs flow selection, data handoffs, and submit commands across several canonical OSMO workflow YAMLs. Day 0 flows cover texture defects (usd2roi ROI cropping, Qwen Image-Edit augmentation, fine-tune-or-passthrough, and inline AnomalyGen labeling including missing-component detection), good/clean-image generation, and structural pose defects (shift, tombstone, sideflip). Day 1 flows perform inference and labeling on real images via real-photo alignment (the default) or manual ROI masks for metal and glass. It ships use-case cookbooks, setup configs (PCB, metal, glass, pretrained), fine-tuning and GPU-sizing references, preflight scripts for URLs, credentials, and pod templates, and OSMO monitoring and troubleshooting guidance.

Target users are manufacturing and physical-AI engineers building AOI defect-detection datasets and models who need first-time asset setup, fine-tuning checkpoints, and deployment configuration without hand-assembling each pipeline stage. It is oriented toward orchestration; individual component internals live in each component's own documentation, and the skill focuses on chaining steps non-interactively and correctly routing vague requests to the right flow.