Surface defect detection is a critical quality control step in metal casting, influencing the aesthetic appeal and mechanical integrity of cast components. Traditional Automated Optical Inspection systems face challenges in detecting defects on non-planar surfaces due to shadows, reflections, and complex geometries. The generalisation ability of existing solutions to new parts or imaging conditions is poor. Existing Vision Foundation models, such as DinoV3 or SAM are only trained on general data on the internet, and perform poorly on industrial data. The objective of this project is to systematically evaluate and compare the performance of different vision foundation models on metal surface defect data. The aim is to test the efficacy of different fine-tuning and adapter strategies for domain adaptation on complex textured metal parts with a variety of surface defects, leveraging the available Metal Surface Defect Dataset (MSDD) and other open datasets of similar type (e.g. GC10-DET), as well as our own data, which has been collected in two ongoing projects at RISE.