An optimization approach that bridges the gap between training and inference techniques via a highly detailed taxonomy of data characteristics to explicitly control generation attributes and implicitly condition generations during inference.