Generative AI algorithms like Stable Diffusion generate images based on the input of a textual description called a “prompt”. Additional image input such as edges, depth maps or sketches allow us to have more control over the output.
But how much influence do we actually have on the result? Are the results neutral and unbiased? The installation Unstable Mirror allows participants to explore the underlying model of the generative process in an explorative and playful way, discovering possible biases of the AI model on their own. The input given to the installation is a picture that is captured when the red button is pressed. The program now extracts the edges from the picture. At the same time, an image recognition algorithm generates a textual description of the image. Based on this input an image is generated using Stable Diffusion.
From this point, recursion starts and Stable Diffusion feeds itself only with the self-generated output. Slowly the original image is alienated, contours and shapes are reinterpreted and variations of the real (spatial) situation are created. Through the recursion, effects caused by possible biases in the model are amplified and become visible.
Programming and AI Pipeline
Hardware Product Design
Video and Photo Documentation
Unstable Mirror is an outcome of the AI+D Lab residency program at the HfG Schwäbisch Gmünd.