Regardless of training data, it isn't matching to anything it's found and squigglying shit up or whatever was implied. Diffusion models are trained to iteratively convert noise into an image based on text and the current iteration's features. This is why they take multiple runs and also they do that thing where the image generation sort of transforms over multiple steps from a decreasingly undifferentiated soup of shape and color. My point was that they aren't doing some search across the web, either externally or via internal storage of scraped training data, to "match" your prompt to something. They are iterating from a start of static noise through multiple passes to a "finished" image, where each pass's transformation of the image components is a complex and dynamic probabilistic function built from, but not directly mapping to in any way we'd consider it, the training data.

It's not the "where" specifically I'm correcting, it's the "when." The model is trained, then the query is run against the trained model. The query doesn't involve any kind of internet search.