Wistron Integrates NVIDIA Cosmos for Generative AI in Smart Manufacturing

Wistron is enhancing its production lines by collaborating with NVIDIA, integrating the latter's "Cosmos" world model into its SMT (Surface Mount Technology) smart manufacturing process. This move leverages generative AI to automatically create defect samples, addressing a critical shortage of abnormal image training data that has long plagued the electronics manufacturing sector. The time required to generate these samples has been reduced to mere seconds, signifying the formal transition of generative AI from proof-of-concept (POC) to actual production line application.

The primary breakthrough in this verification lies in using generative AI to replace the previous method of manually creating defect samples. By lightly fine-tuning NVIDIA's Cosmos model, the system can rapidly generate highly realistic defect samples directly from normal product images. These samples can be adjusted according to different materials, process parameters, and quality specifications, greatly improving the efficiency of data preparation before AI model training.

Wistron requires only a small number of real production line samples to calibrate the model, with the entire process completed within an hour. The generated defect images closely resemble actual production line conditions and can be directly applied to AI visual model training, reducing the human and time costs previously associated with building defect databases.

This verification not only demonstrates the application potential of generative AI in smart manufacturing but also establishes a replicable and scalable AI data generation process. In the future, when facing product design changes, new customer integrations, or novel defect scenarios, the R&D team can quickly update models, enhancing the ability of AI quality inspection systems to synchronize with production lines. Wistron plans to deeply integrate this technology with its existing AOI (Automated Optical Inspection) processes, connecting image acquisition, defect generation, model training, and online inspection to establish a complete AI data loop, continuously improving the efficiency of smart manufacturing and quality management.