Foundation models (FMs) are deep learning models distinguished by their extensive training data and adaptability to a variety of medical tasks including disease classification, monitoring, risk stratification and treatment planning. They are categorized into large language models (LLMs) and vision-centric FMs, which process only text or only images as inputs, respectively. Alternatively, some models integrate multiple modalities, such as images and texts, in vision language models (VLMs), or may expand this multimodality by integrating audio, video, genomics and patient metadata in large multimodal models (LMMs). FMs are expected to help dermatologists in the clinical setting by leveraging advanced AI capabilities to standardize diagnosis and more accurately quantify disease severity, personalize treatment planning and ultimately improve patient outcomes. In this narrative review, we present an overview of the main milestones in generative AI that have driven the evolution of dermatology-focused FMs, a field still under active research. Additionally, we summarize the current landscape and the principal medical FMs that have been developed for image-based medical specialties. Finally, we analyze potential risks and future directions in this field, offering insights from both clinical and technical perspectives.