Mapping the black box. Visual investigation of a diffusion model’s latent space

Abstract

Although the translation from text to images has been a long-standing aspect of human visual expression, generative AI models add a new way to perform these translations based on textual prompts. This new possibility makes the generative models’ internal logic and decision-making processes become central.

The research explores the Midjourney v6.0-mediated translation from text to images through three types of experiments, with a particular focus on the correlation between generated images and specific prompt variations. The proposed methods prove to be a successful strategy to investigate the model’s latent space and decision-making processes, and the analysis of the generated image series reveals intriguing insights about the AI’s ‘black box’ structure and its internal latent representations.