The project Font Hallucinations was created as a semester project in the course "Machine Unlearning" by Nick Schmidt at the Düsseldorf University of Applied Sciences. The project examines an increasingly disappearing phenomenon within generative AI systems: hallucinations – particularly in connection with writing, typography, and semantic readability.
While current AI models are increasingly optimized to produce correct, readable, and controllable results, faulty, ambiguous, or "hallucinated" outputs are increasingly pushed into the background. Font Hallucinations addresses exactly this and understands these artifacts not as errors, but as aesthetic and critical potential. The focus is on the moment when AI attempts to reconstruct writing, but fails and creates new, hybrid forms between image and text.
Creating a simple layout generator. The generator selects from a variety of fonts of the types sans-serif, serif, monospace & display and places randomly chosen words & letters with random font size on a white square background.
A script uses the generated layouts to alienate them through an AI upscaling tool. For the process, the layouts are first downscaled to a low pixel size. Then these downscaled images are enlarged by the AI tool "Upskayl".
Subsequently, the outputs were evaluated as an intermediate step. By experimenting with different pixel sizes in downscaling and playing with font sizes on the layout, values were defined that result in the most interesting effects after upscaling.
After generating some layouts, a second upscaling loop was implemented to obtain even more interesting AI effects and hallucinations.
After the workflow was established, experiments were conducted with various backgrounds and font colors. Colorful backgrounds, monochrome backgrounds with geometric shapes, and image backgrounds were added. The categories were added as options to the generator.
Since the most interesting results were observed with layouts featuring white text on dark background images and black text on light background images, these were also added as options to the generator.
To archive and better evaluate the results, a gallery was created that compares the image generations with their original layouts. Filters allow the different generation options to be viewed separately.