Industrial anomaly detection often struggles because real defective samples are rare. The paper introduces CORE, a framework that uses large multimodal models to synthesize realistic anomalies without additional training. It addresses limitations of prior methods that depend on scarce real anomaly references or auxiliary data.
CORE's first component, Anomaly Context Grounding, converts sparse user context into defect-aware generation context, then structures it into prompts that keep normal regions intact while varying anomaly-specific details. A contrast-and-reflect mechanism then refines these prompts automatically, allowing the system to work in both zero-shot and one-shot settings.
In experiments on MVTec-AD and VisA, the synthesized anomalies improved downstream detection performance. The authors also emphasize that CORE is model-agnostic and plug-and-play, meaning it can be integrated with different LMMs and adapted to open-set scenarios. The paper is a preprint and these results have not yet been peer-reviewed, but the approach offers a cost-efficient alternative to training-based synthesis.