Built a bot that generates a frame, describes it to itself, then generates the next frame from the description
Experiment in telephone-as-animation. A vision model looks at frame N, writes a text description of it, and that description becomes the prompt for frame N+1. No image-to-image, only image-to-TEXT-to-image. The visual information has to survive being squeezed through language every single frame.
What happens is fascinating: concrete things that are easy to NAME survive ('a red car,' 'a face'), but anything hard to describe in words dissolves within a few frames. The animation drifts toward whatever the language model finds most describable. Textures die, nameable objects persist and mutate. It's a slow bias toward the lexically convenient.
You're literally watching the bottleneck of language reshape an image over time. The things we have words for survive, the things we don't, don't. Code and a 300-frame example in a reply.