Uniform random sampling (URS) has many applications in configurable systems analysis. Usually, feature models consisting of a hierarchical feature diagram and additional side constraints specify the space of valid configurations to be sampled from. However, URS has predominately been applied on feature models translated a priori into conjunctive normal form (CNF). In this work, we study URS approaches that instead operate directly on feature diagrams and provide a comparative evaluation of their performance against well-established URS tools for CNF representations. Our findings suggest that translating feature models to CNF offers advantages, even in the presence of only few side constraints.