New image model rendering artist signatures

I am not sure how I feel about image renderings containing the actual signatures of the artists they must be trained upon. I did not specify the rather popular fantasy illustrator "Greg Rothkowski" here. I've noticed the new model is good at rendering words, and I actually think it's doing an overall better job at rendering good results, so congrats to dev on this triumph!

However, this is really pushing the argument for AI a bit into a bad light. If curious, I used the "Fantasy Portrait" style with the following prompt:

the Krovians are a sturdy, ((four-armed)) species with rough, scaly skin. They are known for their resilience and adaptability, having evolved on a planet with extreme temperature variations. Their cybernetic enhancements are often a blend of necessity and aesthetic, a testament to their technological prowess. The driver, with his grizzled appearance and cybernetic arm, seems to be a seasoned individual, likely with a wealth of stories from his time navigating the undercity. The Krovian driver, a grizzled Krovian with a cybernetic arm, looks at you with curiosity before nodding in understanding as you glide over to the open side door...

No negative prompt used.

7 comments

Okay, so since you used a 'Style', it would essentially add additional text/prompt to your input prompt to 'steer' the image to the selected style.
Here is how the prompt is set up using the 'Fantasy Portrait' style on the t2i-framework-plugin generators like ai-text-to-image-generator:
<Your prompt>
, d&d, fantasy, highly detailed, digital painting, artstation, sharp focus, fantasy art, character art, illustration, 8k, art by artgerm and greg rutkowski. It's an absolute world-class masterpiece artwork. It's an aesthetically pleasing artwork with impeccable attention to detail and impressive composition.
So, with the prompt adding 'greg rutkowski' (and artstation) it would add some kind of signature (or watermark) with it. You can view what the styles are adding on this styles import page.
- Yeah, I know styles add a preset list of prompts...It's just that one would think this behavior would be inherently dealt with by now. Part of it's training, so to speak. I guess I'll be more vigilant about describing things without words and signatures, but it's not quite a "style", as in a general kind of inspired creation, if it's including a signature. That's more akin to just "copying." I'm sure this will improve with time.
Flux does not have a negative prompt without witchcraft hacks. I hate flux because of this. You can actually get the flux fine tuning training set as a LoRA on Civitai to apply to a Pony model and get similar results with the ability to remove stuff using the negative prompt.
You must use conceptually opposite prompts to try and remove these elements with flux. Like maybe use unsigned as a keyword.
- Thank you for the reply. I'm a bit less knowledgeable on these tools you've mentioned. Forgive my ignorance, but what is meant by a witchcraft hack?
  Also not sure what I've done to deserve downvotes... Peculiar. Anyways, the "unsigned" and "no signature" advice was quite good. I don't mind simply incorporating negative prompts into prompts; it's a bit more intuitive-- why are there two queues for rendering anyways? I don't order a cheeseburger from a diner twice, one order with everything I want, the other order with everything I don't want (NO ANIME CHEESE, NOT RAW, NOT SMASHED)...
By the by I'm seeing a lot of blatant "Artstation" watermarks and imprintings as well in my results, without requesting this. Maybe the presets should be tweaked to fix this problem.
I'm still using my negative prompt from the previous model, and signatures seem to be rare or minimized. I use multiple artists in my prompts.
extra limbs, extra fingers, too many fingers, deformed, text, (signature, signed, watermark, watermarked signature), picture frame, framed painting, brush, easel, holding paintbrush, bad art, bad composition, terrible lighting, bad anatomy, tattoo

7 comments