The Proof is In the Prompting
So, remember what I said in my previous post about how the images in question illustrated something about the impact things do (and don't) have in an image prompt? Here's an explanation of what I was talking about with that.
Plenty of other people have talked about how some of the common negative prompting that people enter ("bad anatomy, extra fingers, poorly drawn face, ugly, distorted," etc) involves a significant amount of magical thinking- that, because of how most widely available ai art generators are built on top of the snapshot-of-the-whole-internet Common Crawl data and the way those images were labeled, telling the AI to exclude bad hand anatomy or whatnot doesn't reliably get it to not generate wonky hands. The same goes for some common positive prompting, like "trending on artstation".
But, what if I told you that prompting the AI to draw in the style of a living artist is also often just wishful thinking? The image on the left has 'by greg rutkowski' as part of its prompt- Playground automatically appends that (as well as 'trending on artstation, sharp focus, studio photo, intricate details, highly detailed') to the end of your prompt when you create an image using the 'Delicate Detail' style. The image on the right was generated from the same seed, same prompt (including the 'trending on artstation, sharp focus...' part), with the only thing changed being that it was created with no style applied and with 'by greg rutkowski' removed. They're not identical, but they're close enough to be recognizably variations of the same picture, and if they're not in exactly the same style, they're very close.
As a control, this pair of images are from me accidentally running the exact same seed and prompt twice- there's a tiny bit of variation, but it's easy to miss if you're not looking for it and depending on your visual acuity and the size of the screen you're seeing this on, it might not be visible even if you are looking.
And here's the reason why I ended up accidentally duplicating a generation- the image on the right is where I accidentally deleted my negative prompts. Still has the 'by greg rutkowski', though! You can still see that it's a variation of the same picture (/generated from the same seed), but the difference is much greater than with the presence or absence of prompting for a named artist. (Very possibly it could be considered a stylistic difference, but I don't have the technical vocabulary/art school background to be confident about that.)
So, what's the point here? What's the takeaway, what did I learn from these mis-gens?
First, naming an artist in a prompt doesn't necessarily affect the end results all that much. For very widely reproduced/riffed on/repeatedly included artists like Van Gogh, Caravaggio, or any of the other 'Masters' and big influential names you get taught about in art classes, sure, you get a demonstrable difference, but while Greg Rutkowski has potentially valid reason to be upset about people prompting AIs using his name, it doesn't actually do as much as the people prompting probably were thinking it does. Without directly going image-to-image or being specifically trained on a particular artist's works, AI (or at least, Stable Diffusion/SDXL and its descendants) isn't capable of replicating the style of most random living artists, because it doesn't 'know' what their style is in the first place.
Second- those negative prompts may not do exactly what they say that they're doing, but they sure are doing something. Excluding them there may have been an accident, but I might want to noodle around with doing it on purpose sometimes just to see what happens.
