Automated robot-guided welding is a crucial process in automotive body shops. Due to various factors influencing the resulting quality and strict quality requirements, quality assurance remains a critical challenge with several quality control loops. These include among others non-destructive examinations such as ultrasonic testing, which are carried out by domain experts on a small proportion of samples. Because of increasingly available data, there is a trend towards using Deep Learning (DL) for quality control in welding, which has been investigated in research for many years. However, the analysis and comparison of DL architectures to generalize to different process parameters are often neglected. This paper applies state-of-the art DL architectures from the field of Computer Vision (CV) to a real-world dataset from the automotive industry comprising weld samples with different sheet metal combinations and evaluates their generalization ability. The findings contribute to bridging the proof-of-concept to production gap for DL use for welding quality control.