Automated visual inspection systems are commonly used in industry to ensure the product quality. Machine learning approaches show promising results for the automated defect detection using the acquired images from the inspection system. However, real data of defective products are not available in the amount needed to train robust algorithms. Synthetic images similar to those obtained in the real inspection environment can overcome the lack of sufficient training data. We simulate the inspection environment including camera, illumination and the object to be inspected. Therefore, a 3d mesh of the object is used, defect geometries are imprinted and parametric models define the object surface texture. To include the diversity of occurring defects, various defect types have to be modeled. Three basic classes for surface defects exist based on their shape characteristics: point-like structures such as dents, linearly extended defects such as cracks and facetal defects such as delaminations. Here, we focus on modeling surface defects of cast metal objects and restrict on linear and facetal defects. 3d Voronoi tessellations have already been used to generate realistic semi-synthetic 3d images of concrete with crack defects. We adopt that approach to take the poly-crystalline micro-structure of the metal into account. As visual systems inspect the surface and not the internal structures of the object, no detailed 3d defects are required. Complex 2d structures with a simplified 3d component are sufficient. Thus, we use random 2d Voronoi tessellations that form the basis for the stochastic models of both defect classes. Tessellation features such as their edges and cells are used to model different defect types. Moreover, tessellations with different cell intensities model defect structures at different scales. We restrict to parametric models to guarantee the variability within one defect type by using different parameter configurations. Furthermore, applying real parameter values ensures the generation of realistic defects. Thus, a large variety of defect geometries can be produced, which are then added to the 3d mesh of the inspected object to generate synthetic training data in sufficient quantity and variability.