Synthetic Anomaly Generation

Synthetic anomaly generation studies how to create abnormal examples for settings where real defects or rare events are limited, costly, or difficult to label. Common approaches use diffusion and other foundation models for text-, mask-, or reference-guided image editing, alongside procedural simulators, physics-based models, and latent-space pseudo-outlier generation for images, videos, time series, graphs, and 3D data. Synthetic examples support detector training and calibration, localization, and benchmarks with controllable anomaly types and labels. A central tension is balancing realism, diversity, and label accuracy with reliable transfer to real data, since synthetic artifacts or domain mismatch can undermine downstream performance.