As part of my work as professors of data visualization at UNIRSM in San Marino, every year I supervise a few students on their master’s thesis projects. It’s a great opportunity to exchange ideas with students and experiment in new territory. This year I had the pleasure of supervising Arianna Faccioli on her thesis project, “Prompted Identities: Exploring representational bias in text-to-image models”.
In her words: “The project investigates how text-to-image generation models represent people and social roles. Systems such as Stable Diffusion can produce synthetic images from textual descriptions, but what appears to be neutral image generation is actually shaped by technical choices, datasets, and infrastructures that influence what the model is able to represent. The images produced by these systems do not emerge from nothing. Models learn correlations between words and images from the datasets on which they are trained. The composition of these datasets—what they include and what they exclude—determines which visual associations the model learns and which representations it tends to reproduce.”
The thesis resulted in an interactive, informative platform that lets users explore a synthetic dataset of 4,000 images of different professions, generated with Stable Diffusion.
You can explore the project here.
Giorgio.