My humble study on the effects of prompting nonexistent words on CLIP-based diffusion models.

A 2.5-year study reveals 'undictionaries'—nonexistent words that produce consistent, specific effects in CLIP-based diffusion models like Stable Diffusion. The researcher has documented the phenomenon, classified outputs, and created a system for finding more of these hidden prompt triggers. This backdoor exists because users can directly probe the model's latent space. The technique may become obsolete if future models use LLM encoders or add protective layers.