What do you mean a model shift of 5? How do you do that and isn't the default already 5? I tried the ModelSamplingAuraFlow node in Comfy but it had absolutely no effect, and a chatbot told me it's probably because I'm using int8-convrot.

@ferrrett33 Yes that's the correct node, and yes it's subtle, especially at lower precision. I don't know if 5 is the default but it's more noticable on models that are targeting it, like Kromoss. Shift is just about nudging the model away from it's comfort zone, which for most models can result in more detail (up to a point until it breaks). Photography isn't the comfort zone for Moss (we're getting innie pussy that's just a painted-on line right out of the box -- that's not from realistic training data), so I'm finding that over many gens the model shift is lessening the occurance of it's cartoon/semi-real nature. Instead of getting 2-3 cartoon outputs every 10, now I get 1. I'm no ai researcher so maybe it's just confirmation bias because I generate so many images but like you said, it's not hurting anything (unless you don't use shift with a model that wants a very specific amount, like MinimaxH3). Feel free to report back if you find a better value :) Like for Anima, I'm using 1.91 which is essentially just a random value, but it's enough to encourage that model into more variety instead of getting the exact same composition from the same prompt over multiple gens.

@Mas00t Thanks, yes cartoons are a problem in Moss, I got rid of them by prepending something like "A 35mm amateur photograph from the 1990's, sharp focus, high depth of field." followed by the actual prompt. I've been testing a lot of these prepends and each one seems a bit different depending on your word usage, but for Moss you definitely need a prepend like that for some models otherwise they will go half anime or do a bunch of images in cartoons.