Comments
Loading Dream Comments...
You must be logged in to write a comment - Log In
Artist
*Almost*; *nearly*; HEART-BREAKINGLY close to right. This model has some very strange ideas about how doors work, and just won't listen if you try to fix them. I made two very subtle changes to the prompt attempting to stop it hanging two doors in one frame and doing bizarre, magical things with the door opening both ways and sliding across the header or whatever it thought it was doing in those clips. Also added 'multiple door' to my negative prompt.
That did at least accomplish something, though I also believe it's entirely possible that my editing the prompt to make sure the word 'door' only appears in it once had nothing to do with this slight improvement. Simply random variation, which happened to break slightly in my favor. But... there's still two doors in that frame. They're just both swung the same way, and one of them appears like magic out of the wall. I'm charitably ignoring the turn in the wall the prompt said should be straight.
My theory based only on observation of what we can actually see happening when we click buttons with this system, since there's no documentation or explanations of even the basics of how it works, is that the AI model is not, in fact, given the prompt we write directly. There's a parser or pre-processor or whatever the engineers call it in this particular implementation which goes through the prompt and (poorly) analyzes what we wrote, putting it into a defined format the model was trained to understand.
I base that claim on spending a fair amount of time (and compute energy) hitting the "enhance prompt" button. A label, I note, that probably violates the consumer protection laws of multiple jurisdictions, it should more properly be labelled "distort prompt beyond recognition" or maybe "massacre prompt" would be more succinct. My guess (since there's no docs) is that button simply runs the prompt we earnestly labor over trying to get the AI to produce something vaguely resembling our mind's eye image of what we want through the blender of the preprocessor it's going to run before submitting it to the the LLM model.
Though that's an insult to a blender, which at least leaves all the chopped up bits whirling around together even after destroying their integrity and relation to each other. Without exception, every time I run 'enhance prompt', important information is missing from the result. LOTS of it. If I'm even close to right about what's happening, the system is more or less counting on the 'ooo, ahhh' factor of beautiful results for users to go, "well, not what I wanted, but really pretty - might as well make it public and get a few likes". Not a lot of artistic integrity or fidelity to our vision rather than randomness in that process, but I suppose it has some quite valid uses. Just not what at least one paying customer is looking for. <sigh>