>>8884265
this is supposed to depend on the model and how it was trained. 1.5 models generally were trained on images that may have had a few tags or were otherwise labeled certain ways in the LAION dataset. XL Models were a mix of images captioned that way and with longer captions, some generated with AI (Albedobase is my favorite example of this - you can get some excellent images with 5 tags and with long prompts, but they do look very different aesthetically). With Flux, the images are described by an AI, which then becomes the image caption.

While this can theoretically offer the proompters more accuracy, the tradeoff is twofold, the larger one being that AI language models are very verbose and no one wants to type a five paragraph essay into forge to get Asuka pegging Shinzo Abe. The second is that each small letter combination is reduced to a "token" and given the purple prose in AI language models, there are a lot of tokens which don't actually mean anything in terms of content but which become necessary for image quality. This is where a niche arises for a model like Chroma - beyond NSFW, it is basically a Flux model that has had a lot of the "padding tokens" removed and as such it is easier to prompt for.