Incorporating flexibility and dynamics of protein targets is a frontier of computational drug design. A machine learning model that jointly generates ligands and bound conformations of binding pockets holds promise to access a larger chemical space by removing unnecessary structural constraints, and opens the door to design campaigns in which only the unbound structure of the therapeutic target is known.Here we report on progress we made towards this goal and present two models: DrugFlow, a new generative model for structure-based drug design with static protein structures that shows strong performance compared to previous methods, and FlexFlow, an extension of this model that also predicts side chain torsion angles together with preliminary empirical data.