Generative Models for Defect Synthesis is the use of generative AI (GANs, VAEs, diffusion models) to create realistic synthetic defect images — augmenting limited real defect datasets to improve classifier training and address severe class imbalance.

Generative Approaches

• GANs: Conditional GANs generate defect images by type. StyleGAN for high-resolution synthesis.
• VAEs: Variational autoencoders for controlled defect generation with interpretable latent space.
• Diffusion Models: DDPM/stable diffusion for highest-quality defect image generation.
• Cut-Paste: Synthetic insertion of generated defect patches onto normal background images.

Why It Matters

• Class Imbalance: Some defect types have <10 real examples — generative models create hundreds more.
• Privacy: Synthetic data avoids sharing proprietary fab images with external ML teams.
• Rare Events: Generate realistic samples of catastrophic but rare defects for robust training.

Generative Models are the defect image factory — creating realistic synthetic defect data to augment limited real-world samples for better ML training.