SMT Perspectives & Prospects: What I Hope the Next Generation Will Carry Forward
Learning With Leo: Observations from IPC Builds in Paris
Wistron Corporation has successfully deployed NVIDIA Cosmos open world foundation models to power AI-driven quality inspection in its SMT production lines, addressing one of the most persistent bottlenecks in high-end electronics manufacturing — the scarcity of defect training data. The implementation enables defect samples to be generated in seconds, marking a significant step in moving generative AI from proof-of-concept into full-scale industrial deployment.
Generative AI Slashes Sample Preparation from Days to Seconds
At the heart of Wistron’s breakthrough is the replacement of manual defect fabrication with a generative AI pipeline. Traditionally, producing sufficient defect samples by hand could take days — a process that now takes seconds. Now, using NVIDIA Defect Image Generation skill powered by NVIDIA Cosmos and NVIDIA TAO Wistron generates high-fidelity defect images directly from normal production images — on demand and at scale. The workflow can be rapidly adapted across varying materials, process parameters, and quality standards, dramatically accelerating data readiness for AI training workflows.
Using this technology, Wistron built a synthetic data pipeline for surface defect detection that eliminates annotation costs, generates realistic defect samples in approximately 1.15 seconds (vs. hours manually), and improved detection accuracy by 5–20%.
Building a Scalable, Repeatable AI Data Pipeline
Wistron sees this establishing a replicable and scalable framework for AI-driven synthetic data generation in smart manufacturing. As product designs evolve, new customers are onboarded, or novel defect types emerge, engineering teams can rapidly retrain and redeploy models — giving AI quality inspection systems the agility to stay in sync with dynamic production environments.
Synthetic data generation will be available in Wistron's existing Automated Optical Inspection (AOI) workflows by June 2026, creating a closed-loop data pipeline spanning image capture, defect synthesis, model training, and inline inspection. The end-to-end system is designed to continuously strengthen quality control operations and advance Wistron's broader smart manufacturing capabilities.
