Generating high-quality photo-realistic visual content is a thriving area of research. In this talk, I will focus not on the generation process, but on understanding how rich visual semantic information is represented within generative models. Specifically, I will present our work that uses mechanistic interpretability tools like k-sparse autoencoders to probe various layers and denoising timesteps of different diffusion architectures. Next, I will present how to leverage interpretable concepts to control the generation process towards more safer and creative outcomes.