Double update March 3, 2024

@DeViLDoNia looking forward to it, since you're into loras, I thought you might be interested in what I'm doing atm, I'm currently looking into leco loras as a way to increase conceptual understanding and creativity, a capability all models have is that if you write two words "a-word b-word" in the prompt and the opposite order "b-word a-word" in the negative prompt, you get a strong cancellation effect that sort of splits the latentspace, for example, if you write P:"girl cat" NP:"cat girl" you'll force the AI to draw whatever is most "girl cat" while trying to be the least like a "cat girl" (stronger effect at higher scale), if the a-b order is reversed, the reverse is true ofc, this forces the AI to "make up it's mind" in a very brutal way, it's can't just relay on generically drawing the "girl" and "cat" as elements, but rather must focus on what the order really means, for this example, if the model isn't totally stupid, it'll draw a girl and a cat separately one way and a cat girl the other way, if it's smart there will be no cat-ears on the girl with a cat, if it's a less good model it'll not split the latent-space cleanly, some elements will end up on the wrong side of the association space, with a good model, no matter what words you split you get interesting creative results, poor models draw more creative "blanks" basically, where as those with a deeper understand of how things relate to other things really shine, knowing all this has made me think that maybe training conceptual pairs is they way to go, where the order really matters, as a clean way to improve models, ideally, the two orders should be as different as possible, some example "being angelic" vs. an "angelic being" or a "forest fantasy" (perhaps going camping) vs. a "fantasy forest" (glowing magic mushrooms parhaps), you could even solve specific association problems with this, for example the word "suit" is very poisoned, this is because the entire western world uses the standard b&w suit, but suit is actually a dual meaning word, we have swim suit, mechanical suit, etc where suit means "x-themed often tightly around the body", the other order has to do with sub-typing, so "suit victorian" would be a latentspace transformation from a normal suit to a subtype of suit, this is the noncreative order and different conceptually from themed suits, if you can find that perfect "partner-word" to something problematic like "suit" you can fix the concept on a very deep level, my try would be "lovecraft", this is because a lovecraft suit would/should be a horror-themed suit (we're generalizing from swim suit etc), where as a suit lovecraft should be very similar to a "suit victorian"/"victorain suit" since the lovecraft setting contains this duality between horror and posh upper-class (unlike victorian itself which is very one-note), my theory is that if you train keyt read orders at the same time, the model can't simply "dump" it's understanding of horror or fancy-ness directly into the tokens "lovecraft" or "suit", it's forced to place the difference into the order itself or the "connection", then when you're prompting the AI, it'll constantly be on the lookout for what type of understanding is most suitable in the current context, rather then simply going the easy way of memorizing what a cat looks like and going "oh, a cat I know what that is", anyway, hope you found this interesting and good luck with the lora training!