Conditional variational autoencoder (cVAE) are used throughout the literature in the sense of a performance-based generative design. One main advantage is the possibility of an inverse problem formulation, allowing the exploration of design possibilities/variations on a pareto-font. This allows for a more dynamic design, including performance and code based parameters as a preselection criterion for the overall design. The work in the presented paper takes this general idea of cVAEs and applies it to connection design, e.g., welded knee connections from I-shaped sections. All created data sets used for the training of the models are based on component-based finite element simulations (CBFEM), performed with the software IDEA StatiCa Connection and the built-in python API. The overall focus of this paper is specifically set on data creation/post-processing, data reduction, and its manipulation. Furthermore, different cVAE model architectures, as well as their hyperparameters, are compared in order to make recommendations on the training performance and the prediction accuracy. Subsequently, this paper shall demonstrate a dynamic design process, using cVAE-based models as a co-pilot to navigate through feasible connection design options.