Developer Alex Van Doren has put a diffusion image model inside a PDF. The document performs the generation itself rather than sending the prompt to an image-generation API, then displays the result as text arranged into a small drawing.
The model was trained on Google’s Quick, Draw! doodle data and generates drawings from 131 object categories. It produces its picture through repeated denoising, a process that gradually turns noise into an image. PDF form fields display the result as 28 rows of text instead of a conventional image canvas.
This is a tightly constrained experiment. Words outside its object vocabulary get mapped to a known category, often poorly, according to the developer. It is not a general-purpose image generator.
The PDF requires Chrome or another Chromium-based browser; Firefox and Safari are unsupported. The project builds on Evan Zhou’s llm.pdf, which demonstrated a language model running within the same document format.