Chloe McFadden
🌟 Work
🌟 Contact
Prompt shifting: Dinner (2024)
Prompt shifting is a creative practice or process I am exploring and developing in my practice-based research.
Prompt-shifting is a method that engages prompt-weighting and img2img models to probe how learnt patterns and biases of genAI applications manifest visually in strange, subtle and distributed ways. Prompt weighting is positioned as a method which allows users to emphasize or de-emphasize certain parts of the prompt resulting in greater control of an image. Prompt-shifting then, is a practice of slowing shifting the emphasis of certain subjects and aesthetics within images to explore and expose the ways in which the learnt patterns associated with specific semantic terms manifest visually.
In December 2024, I hosted a workshop designed to disrupt metaphors which conflate our imaginations with AI outputs. This workshop was held for 8 artists and consisted of different techniques to reveal the politics and biases of genAI images and time for the artists to collaborate and discuss what it means to imagine individually, collectively and as mediated by algorithms.
Featured in the above image is one demonstrated exercise in which by subtly increasing the weights of descriptive terms like “female” or “male” in the prompt, the social and aesthetic encodings of these terms can be exposed and critiqued. As the term “female” gains greater influence, our dinner scene slowly diminishes until we are left with a plate of flowers.
By tracking and tracing what changes (e.g. size, colour, shape, object, tone), why these changes may be occurring (e.g. the dataset, stereotypes, biases) and how these changes alter our interpretation of the images, we can begin to sense and trace some of the temporary ripples and ridges of the ‘map’ of genAI activated via our seed, prompt and image. In this, I argue that prompt-shifting can be another way of ‘reading’ AI images and demonstrating how semantic terms may bias outputs in ways that are unintended and difficult to detect.