Visual defect detection is an important aspect of quality inspection. High performance manufacturing can benefit from automating defect detection, and deep learning techniques are the current state of the art for this computer vision task. The premise of deep learning is that an over parameterized model learns to generalize in performing a task such as classification or object detection by exposure to a wide variety of training examples. In the canonical example, a classifier that has seen thousands of pictures of dogs and cats in different situations and backgrounds will be able to generalize to tell one animal from the other in a newly obtained photo. The assumption that the training data contains great variety is not met in typical manufacturing data sets. Data is highly repetitive and mostly represents defect-free products, meaning there are few images of defects or deviations to learn from. In this training regime, deep learning models are easily over-fit to the training data and can fail to detect defects in the face of variations such as position, lighting, or data drift. In this work, we explore training defect detection models with images of defects on different backgrounds and in different locations, in order to approximate the exposure to highly diverse data sets that is an assumption of a well-trained deep learning model. We demonstrate how models trained on diverse images containing a common defect type can pick defects out in new circumstances. Such generic models could be more robust to new defects not found in data collected for training, and can reduce data collection impediments to implementing visual inspection on production lines. Additionally, we demonstrate that object detection models trained to predict a label and bounding box outperform classifiers that predict a label only on held out test data typical of manufacturing inspection tasks. Finally, we studied the factors that affect generalization in order to train models that work under a wider range of conditions.