Designing a small-molecule therapeutic is a challenging multiparameter optimization problem. Key properties, such as potency, selectivity, bioavailability, and safety, must be jointly optimized to deliver an effective clinical candidate. We present COATI-LDM, a novel application of latent diffusion models for the conditional generation of property-optimized, drug-like small molecules. Diffuse generation of latent molecular encodings, rather than direct diffusive generation of molecular structures, offers an attractive way to handle the small and mismatched datasets that are common for molecular properties. We compare different diffusion guidance schemes and sampling methods using a pre-trained autoregressive transformer and genetic algorithms to evaluate control over potency, expert preference, and various physicochemical properties. We show that conditional diffusion enables control over the properties of generated molecules, with practical and performance advantages over competing methods. We also apply the recently introduced idea of ​​particle guidance to enhance sample diversity. We prospectively survey a panel of medicinal chemists and determine that we can conditionally generate molecules that match their preferences via a learned preference score. Finally, we present a partial diffusion method for local optimization of molecular properties starting from a seed molecule. Conditional generation of small molecules using latent diffusion models on molecular encodings offers a highly practical and flexible alternative to previous molecular generation schemes.