Successful automation of surface defect recognition using magnetic particle inspection (MPI) or fluorescent penetrant inspection (FPI) requires AI models to be trained on high amount of high-quality data. However, situations where non-conforming parts or their usable digitized records for successful AI training are not available are unfortunately quite common (more than 50% of the cases). The lack of data is caused by many reasons (most often a low percentage of non-conforming parts, non-existing parts before the production phase, non-existence or untraceability of previous indication records, low quality data and its unusability, etc.). There are several ways to proceed in such cases. In the case of a low "scrap rate", different approach like "Anomaly detection" method is more suitable because of the possibility to train the AI ​​model only on OK parts. However, this solution has a number of limitations and typically brings a higher "False Positivity" of outputs than can be accepted. Other methods include collection of digitized data directly at the production and inspection site, or creating physical defects, their indications and digitized records mechanically. While these methods might bring positive results there is another easier way that enables production of high amount of high-quality data of any type of defect indication on any type of part – rule based synthetic data generation. If you want to learn more about the pitfalls of these paths, or what the entire training data set usually looks like, you are invited to a discussion, we will be happy to share our experiences.