Generative AI and predictive AI are increasingly transforming quality control by enabling the creation of realistic synthetic defect datasets and training high-performance models that can serve as baselines for real-world manufacturing inspection systems. This study presents a two-stage AI framework for industrial quality control that integrates synthetic data generation via Blender’s Python API with predictive modeling using a custom-built convolutional neural network (CNN). Procedural generation was employed to create a high-fidelity dataset of 2,250 bottle images evenly distributed across three defect categories—no defect, missing cap, and misaligned cap—offering precise control over geometry, lighting, materials, and defect parameters. This approach addresses the challenge of limited real-world defect data by producing physically accurate, reproducible images that closely mimic production conditions while enabling systematic variation of defect attributes. The optimized CNN architecture was designed to balance computational efficiency with high classification accuracy. Experimental results demonstrated that the custom CNN achieved over 99% accuracy, with perfect classification in two categories and minimal misclassifications in the third. Precision, recall, and F1 scores remained consistently high across all classes, and early stopping ensured robust generalization without overfitting. The findings confirm that combining controlled synthetic data generation with a task-specific CNN yields a scalable, accurate, and reliable defect detection system suitable for Industry 4.0 manufacturing environments.