An interactive explorable on the assumptions text-to-image models make before you say anything: where they live, when they lock in, and what it costs to override them.
Measured on Stable Diffusion 2.1 in depth, and replicated on six further models. 54 frozen prompts × 50 seeds × 7 models, with every number on the page traceable to the analysis that produced it.
Explainable AI × human–computer interaction · thesis explorable.
See HOW_TO_START.txt to run it locally, ATTRIBUTIONS.md for credits.