Donmai

AI-generated art check thread

Posted under General

8253803 said:

Oh for the sake of all that's holy. By 2024 we'll get to "you either stream your painting process on YouTube and publish the VOD or you're using AI" level of witchhunting. How can an artist get a timelapse out of an already finished work?

https://twitter.com/amakidaisuke/status/1670855936532164608

This latest screen recording is quite convincing to me. It has all the brushstrokes and rough edges of a human-created PSD. The shading is done layer-by-layer. The ordering of the layers is logical. if anyone can name an AI which generates PSD that looks like this, I might change my mind, but for now I believe it's fully human.

That's the same video they deleted earlier and I was trying to put into words, don't ask me, I laughed at the timelapse comment too. Hive was indeed the basis for this stuff on Twitter. But anyway, the post has been approved.

EDIT:

Maiden_in_Orange said:

Edit: Sorry Individual. I wasn't sure if you knew, and that wasn't my place to talk in the first place.

It's fine.

Updated

Marlor said:

Bruh, this guy may be right but he's so rude.

This guy needs to publish a paper on some reputable machine learning journal about the reliability of this AI detection tool they trust more than their own eyes, before they're worth taking seriously. I've just seen with my own eyes how unreliable tools like this are:

8253803 said:

I tested it using a PSD from an artist's Patreon, exporting two images with backgrounds on / off, and noticed that this tool spits out false positive when the image has a background? Yeah gotta be careful when using this tool.

background on: post #6404534, 73.8%
background off: https://files.catbox.moe/8m4m2r.png, 3.1%

Marlor said:

Why using 3+ colors on 1 layer? That's pretty uncomfortable

Did they? The layers are all organized as flat colors first, and then highlights and shadows on top of that. Smaller components might be done in one layer because they're simple, but skin and clothes are both layered in a way that a human would.

BobTheBuilder_v1 said:

https://danbooru.donmai.us/media_assets/13005637
https://danbooru.donmai.us/media_assets/13007487

Artist talked about using Stable Diffusion for the background in their previous work https://www.pixiv.net/en/artworks/109071720 , but also mentioned in the comment section of https://www.pixiv.net/en/artworks/109168028 that the characters were rendered using Blender and the girl's body had to be fixed with a JPEG in post due to Blender problems.

VR-Man said:

The guy himself posted this on his Twitter.
(And also check out his other tweets)

*Saw this message on danbooru
You know anon, I purposefully didnt try to make it look human because frankly I dont care. The Fn fal piece I make no one recognize it as AI, and for you to come here and to be like "oh watch out is AI" seem a bit clowney to me.

Did they also "purposefully" leave the incorrect number of fingers as-is? Like, the easiest mistake to spot and the most trivial to correct? LMFAO.

1 33 34 35 36 37 38 39 40 41 118