Iterative Visual Interaction with Latent Diffusion Models

Image synthesis generative Artificial Intelligence has the potential to revolutionize many industries, from rapid prototyping of consumer goods’ design to creating works of art. However, the most widespread of these models come in the form of text-to-image models. Controlling their output is often a difficult and imprecise process; indeed, slightly different prompts can lead to significantly different outcomes. Therefore, we develop a novel way to interact with generative AI that mimics the way in which humans naturally create works of art and design. We achieve this more granular and intuitive interaction method by exploiting the latent space of Latent Diffusion Models to generate image variations.