Sometimes you get a composition from the specific prompt+seed+etc. And it has an alien element that is surprisingly stable and consistent throughout “settings wiggling”. I guess it just happens that training may smear some ideas across some cross-sections of the “latent space”, which may not be that explicit in the datasets. It’s a hyper-dimensional space after all and not everything that it contains can be perceived verbatim in the training set (hence the generation capabilities, afaiu). A similar thing can be seen in sd lora training, where you get some “in” idea to be transformed into something different, often barely interepretable, but still stable in generations. While you clearly see the input data and know that there’s no X there, you sort of understand what the precursor is after a few captioning/training sessions and the general experience. (How much you can sense in the “AI” routine but cannot clearly express is another big topic. I sort of like this peeking into the “latent unknown” which skips the language and sort of explodes in a mute vague understanding of things you’ll never fully articulate. As if you hit the limits of language and that is constraining. I wonder what would happen if we broke through this natural barrier somehow and became more LLMy rather than the opposite). /sot