Conditional Generative Models (CGMs) are powerful tools for editing complex 2D or 3D content with high quality, and have been particularly useful for crafting visual content involving humans. However, many existing CGMs still impose a number of constraints on training data, which limits their flexibility as a tool for generating novel content in practical settings. This thesis makes a step towards improved flexibility and quality of CGMs in the context of human face and body editing. Firstly, this thesis addresses the reliance of prior conditional generative adversarial networks (CGANs) on labelled training data in image-to-image translation (I2I-T) tasks. It proposes a self-supervised learning (SSL) constraint for CGANs, utilising unlabelled data more effectively and drastically reducing the labelled data required without impacting quality. Secondly, the thesis tackles the lack of semantic understanding in label-guided CGANs for facial attribute transfer. Whilst prior methods proposed using user-provided target semantics as guidance, this thesis proposes a parallel CGAN framework which learns to generate this target information on the fly, improving editing quality and preventing erroneous edits. Lastly, the thesis reduces the reliance of prior 3D Pose Transfer methods on ground truth output and correspondence for training. It proposes a solution which refines the model's latent codes and incorporates SSL techniques to improve performance and enable unsupervised learning, achieving state-of-the-art results and improving model generalisation.