cspicyweener said:
THE FOLLOWING IS NOT AN ACCUSATION!

I think I see where the 'AI' conflict may be coming from.
It's like the image was, in fact, properly drawn.
THEN an AI was used for the 'clean up'.
The image looks 'too clean', if that makes any sense.
And to that end, if EVERYTHING was done by hand, then good work.

I completely disagree with the idea that an image can be "too clean", considering tags like binary_drawing consist entirely of perfectly-clean illustrations.

ca7m0use said:
I can tell at a glance what's AI and this isn't it, not Shure why they keep getting deleted but this is really good keep it up

No, you can't.

kristiqn_951 said:
Don't stop drawing the way you do it's 10/10.
Just put the ai poison pattern so the ai detection fails but I don't actually know if it works

Our AI-generated deletions are manual decisions made by real humans. Adding adversarial noise is not magically going to allow AI-generated artwork to approval, and I'm not sure what decision-making process would have you think that we use AI to detect AI.

post #922505

Threezetas's s contain:

Most of these are at least occasionally identifiable with an amateur digital artist who's finding their sea legs, but the blush sticker pattern elevates this from "well-meaning but suspicious artist" to "overt fraud", and their commenting history more or less reinforces this by misdirecting in ways like "maybe the janitors thought (X)" or defaulting to the classic excuse "it's my style".

Please compare the two images below.

https://twitter.com/thebird444/status/1855629399896330283
https://twitter.com/thebird444/status/1856013156855767427

These are two different stages of the same image. In spite of having linework for a Lucario, notice that the colored version mysteriously introduces the omnipresent generative blush stickers. In a manual coloring workflow, this would simply never occur. There is no magical brush or stamp that produces these slightly irregular multi-layered blush stickers, this exact style of blush is simply an artifact of how programs like Stable Diffusion resize and refine their outputs. If there were a stamp or a brush capable of generating these kinds of blush stickers, then one would expect every one of Threezetas's blush stickers to look perfectly similar. They're not. Jolteon's blush is not perfectly similar to Umbreon's, Eevee's are not perfectly similar to Flareon's, etc.

The simple explanation (and therefore the one that's most likely to be correct, taking into Occam's Razor) is that Threezetas is generating Pokémon, drawing linework based on the output, and then combining manual coloring and AI-generated coloring to create the final product. Our current standards do not allow incomplete paintovers like this to approval, and so all of these deletions are correct.

If Threezetas takes umbrage with these accusations then they need to conclusively prove that all 100% pixels are manual art: go the ai generated reference route and release the reference image separately, demonstrate their workflow from start to finish by means such as streaming, distribute their layered source files for examination, things like those. I will not accept anything less than a direct response to the accusations levied by the other post approvers.

Updated

  • |
  • 4