Training Smarter Defect Detection —
Without Waiting for Defects

The Data Problem That Stops Most Teams Before They Start

Building a reliable AI model for visual defect inspection has always come down to one hard constraint: defect data is scarce. In manufacturing environments, defects are rare by design — which is exactly what makes collecting enough labelled examples such a slow, expensive process.

Teams either wait months for defects to accumulate in production, or they ship models trained on too little data and accept the accuracy trade-off. Neither is acceptable when inspection quality directly affects product yield and customer outcomes.

At CODETRACE, we addressed this challenge not by collecting more data, but by generating it with precision. At NVIDIA GTC Taipei 2026, we are showcasing how we integrated the NVIDIA Defect Image Generation skill — a few-shot, diffusion-based defect generation pipeline built on NVIDIA Cosmos, with NV-DINOv2, C-RADIOv3, and SAM2 powering the visual encoding, quality evaluation, and mask refinement stages — into our Automated Optical Inspection (AOI) pipeline. Starting from a small set of real annotated defect images, the pipeline synthesises a large, diverse, and pseudo-labelled training set, which we then use to train a production-ready inspection model — compressing what was previously a months-long data-collection cycle into days.

Why Standard Approaches Fall Short

Modern industrial inspection increasingly relies on AI models trained to detect rare, high-impact defects. But in practice, those defects don’t show up often enough to build a robust training dataset within a reasonable product timeline.

Traditional synthetic data generation (SDG) approaches have tried to address this, but they typically struggle with realism — generated defects can look artificial, causing models trained on them to underperform when confronted with actual production defects. Physical consistency and spatial accuracy are hard to preserve without grounding the synthesis in real imagery.

This is precisely the gap that NVIDIA’s Defect Image Generation skill — part of NVIDIA’s Physical AI agent skills for Synthetic Data Generation — was built to close.

Where We Started: Direct Training on Real Images

Our in-house vision model, Codetrace DetecXion, is CODETRACE’s purpose-built engine for automated defect detection. Before integrating AnomalyGen, our standard pipeline was straightforward: collect real defect images, annotate them, post-train DetecXion directly, and deploy.

With only 10 real annotated images as our starting point, this baseline approach gave us a deployable model — but one with limited ability to generalise to defect variations it hadn’t seen in training. It was the fastest path to deployment, and it served as our reference point for measuring what synthetic augmentation could actually deliver.

Baseline Workflow — No Synthetic Data

Direct training on 10 real images · 5 stages

Baseline

Scaling to 200 Images — Without a Single Additional Real Defect

Integrating NVIDIA’s Defect Image Generation skill required rethinking the data preparation phase entirely. Rather than feeding our 10 real images directly into DetecXion, we first put them through a diffusion-based inpainting workflow — Codetrace AnomalyGen orchestrating Cosmos-Predict2, SAM2, and Cosmos-Reason1 — which expands the dataset 20× before post-training in NVIDIA TAO Toolkit.

The Defect Image Generation skill is part of NVIDIA’s Physical AI agent skills for synthetic data generation, built on NVIDIA Cosmos. For a full description of the skill, see NVIDIA Cosmos.

The enhanced eight-stage pipeline works as follows:

End-to-End Pipeline — NVIDIA Defect Image Generation Skill

Real data → Synthetic augmentation → Train & deploy · 8 stages

+ AnomalyGen

Stages 1–2 remain unchanged from the baseline: annotate the 10 real defect images using Codetrace LabelMe and set up the Docker / CONDA / CUDA environment.

Stage 3 is where the pipeline diverges significantly. Instead of heading straight to model training, we fine-tune NVIDIA Cosmos-Predict2 (using LoRA) on our 10 annotated real images — teaching the model the visual, textural, and material distribution of the target defect type. NV-DINOv2 and C-RADIOv3 power the visual encoding and quality evaluation at this stage.

Stage 4 uses SAM2 to automatically identify and place regions of interest (ROIs) — the precise spatial locations on each image where defects can realistically appear. This ensures that synthetic defects are placed with physical plausibility, not at random.

Stage 5 is the core: the NVIDIA AnomalyGen skill leverages the trained diffusion model and SAM2 masks to generate photorealistic synthetic defect images grounded in the real image’s structure, lighting, and material characteristics — supporting PCB, metal surface, and integrated circuit device defect types. This is where our 10 images become 200.

Stage 6 runs Cosmos-Transfer1 to apply Sim2Real style transfer and lighting normalisation to align synthetic images with real factory imaging conditions, then Cosmos evaluation scripts validate image fidelity before training. Stage 7 runs SAM2 + Cosmos-Reason1 to automatically pseudo-label all 200 outputs — SAM2 produces defect masks and Cosmos-Reason1 provides semantic captions for class assignment, eliminating manual labelling overhead entirely.

Stages 7–8 bring it home: Codetrace DetecXion is post-trained on the full expanded dataset of 200 images using NVIDIA TAO, and deployed to production as a fully inference-ready model.

Baseline Pipeline vs. With Codetrace AnomalyGen

Measured Accuracy Results

Validated against a 100-sample running set (70 genuine good, 30 genuine NG).

What This Means for Industrial AI Development

This integration demonstrates something important: the bottleneck in production AI for visual inspection is rarely the model architecture — it’s the data. Codetrace AnomalyGen tackles this directly: Cosmos-Predict2 and SAM2 inpaint defects into real product images, and Cosmos-Reason1 validates every synthetic sample before it enters training — yielding a 20× dataset expansion grounded in genuine factory imagery.

“The real breakthrough is not generating more data — it is generating the right data. NVIDIA’s Defect Image Generation skill gave us synthetic defects that are physically grounded, spatially accurate, and realistic enough to meaningfully improve a production model.”

— Mr. Jason Ng, CEO of CODETRACE Sdn Bhd

Expanding Across Defect Types and Materials

The current pipeline has been validated across PCB, metal surface, and integrated circuit device defect types using the Codetrace AnomalyGen + NVIDIA Cosmos pipeline. CODETRACE will continue expanding material coverage and integrating Cosmos-Transfer1 for Sim2Real scenarios where real reference imagery is unavailable. Trained models are deployed to factory cameras through CODETRACE’s own inspection platform, Codetrace DetecXion.

For teams facing the same data scarcity challenge in their own inspection workflows, this approach represents a practical, production-tested path forward — one that eliminates the most common reason inspection AI projects stall before they ever reach deployment.

For manufacturing teams confronting the same data-scarcity challenge, this approach represents a practical, production-validated path to Day-0 inspection readiness. The full workflow — Codetrace AnomalyGen orchestrating NVIDIA Cosmos-Predict2, SAM2, Cosmos-Reason1, and TAO Toolkit, with deployment via Codetrace DetecXion — is available to manufacturers as a turnkey integration. To explore how NVIDIA Cosmos synthetic data generation and Codetrace DetecXion can accelerate your inspection AI programme, contact our team.

📍 Visit us at NVIDIA Inception Pavilion at InnoVEX 2026 — Booth S0712a, 4F, TaiNEX Hall 2, Taipei — 2 June to 4 June 2026