unStable Mirror
An AI mirror that re-renders your reflection through a generative model, over and over, until only the machine's assumptions look back.
unStable Mirror lets you watch an AI watch you. Step in front of the custom-built mirror frame, press the red buzzer, and the integrated camera captures your image — then hands it to a generative model, again and again, until what looks back is no longer you but the model’s idea of you.
Each captured image is reduced to two machine-readable descriptions: an edge map, extracted with canny edge detection, and a caption generated by an image-classification model (CLIP ViT-L/14). Both are fed into Stable Diffusion, which synthesizes a new image from them — and that result is fed straight back into the pipeline.
The recursion is the piece. With each cycle, contours are reinterpreted, shapes drift, and whatever the model assumes gets reinforced. Small algorithmic preferences that would stay invisible in a single generation amplify iteration by iteration, until the mirror reflects less of your appearance and more of the assumptions, biases, and aesthetics baked into the model.
The installation is built to make this drift experienceable for a general audience — no explanation required, just a button, a familiar mirror situation, and a transformation you can follow with your own face as the test subject. It is a conversation starter about what generative systems actually do: not reproduce reality, but re-render it through learned expectations.
unStable Mirror is an outcome of my AI+D Lab residency at HfG Schwäbisch Gmünd; many of its ideas later consolidated into Transferscope. Thanks to Aeneas Stankowski, Alexa Steinbrück, Rahel Flechtner, and Benedikt Groß for their support during the residency.
