Training deep learning models for computer vision tasks requires large volumes of training data. Collecting such data is not difficult, but privacy regulations restrict the use of personal data without getting explicit consent or rendering the data anonymous first. Traditional techniques for anonymization, such as face blur, negatively impact the utility of the data in training. A more sophisticated approach is realistic anonymization, i.e., the replacement of identifiable data with synthesized realistic data. This thesis aims to investigate how realistic anonymization can be applied to images of people, the degree of privacy it ensures, and what effect this has on the utility of the data. Specifically, the diffusion-based image anonymization pipeline RAD is introduced and evaluated on privacy and utility.