A new open-source framework lets researchers manufacture fake 3D defects on demand — the better to catch real ones.
AnomalyFactory 3D (AF3AD) is a modular synthesis tool designed to create artificial defects in 3D point cloud data. Because real defective parts are rare and expensive to collect, most anomaly detection systems are trained only on normal samples, which limits how well they generalize. AF3D fills that gap by warping normal geometry into plausible-looking flaws using a deformation model built around local geometry frames, with controls for spatial falloff, directional bias, and surface displacement. The result is a library of geometric defect presets that can be dropped into existing detection pipelines. The researchers tested it with two different detector types — one that predicts offsets, one that reconstructs surfaces — and saw consistent gains on the AnomalyShapeNet and Real3D-AD benchmarks at both the object and point level.
Manufacturing and robotics inspection depend heavily on catching defects in scanned parts, and the scarcity of real anomaly data is a known bottleneck. A standalone synthesis tool that transfers across detection paradigms without retraining from scratch could lower the barrier for teams that cannot afford to wait for a defect-rich dataset to accumulate.
The code is public on GitHub, which is the right move for academic tooling — though how well AF3AD generalizes beyond benchmark datasets to messy real-world scans is a question the paper, by design, leaves open.