@hideki I think that there is merit to AI art, but it's a different kind of merit. The analogy I'd use is that one is like running and the other is like driving. The goal of both are the same, but the skillset required to be good at both are very different. The runner could say that driving is easy because they aren't using their legs, but there is a lot that goes into maintaining and optimising cars for certain purposes and, say, using that car to race.

@w

I don't know, probably there is some kind of merit, but i seriously don't think the amount of skill merit is by any stretch similar.

You could argue that yeah, having a car in a good state enough to be in a race has merit (even if you just bought the parts and someone else installed them), but then again, you can't really claim that you should be given the gold medal in 100 meters flat because you ran through it faster than other people that trained every day using their legs.

Then you have these people pumping AI imagery nonstop at a breakneck speed and polluting spaces that traditionally were for people to exhibit their skill. Pixiv is having that problem.

@hideki I agree that they should be clearly marked, and not directly compared to human art. That's kind of my point, they require different skillsets.

Recently I've been trying to teach a model how to make images of Webber, a character from an indie game with poorly labelled art which mostly lives outside of boorus. There were a lot of steps and complications in making it.

1. I had to actually acquire the images. Most of the art of Webber that exists is not on booru sites. Luckily I had a scrape from DeviantArt already that had about a hundred and seventy images that used a custom program I wrote. I had to go through each and rule out any that didn't fit what I needed for the model, which only got me around twenty images or so on top of the existing twenty or so I source from Klei art (which also had rejects and major edits I needed to do.)

2. I had to crop them to a 500x500 square, which might not be a convenient size for the art. I had to edit multiple images to fit properly.

3. I had to ensure the artist didn't add any clothes or items that might confuse the AI. There are images in the dataset I manually edited things out of (such as one image where I removed a chain Webber was wearing with an item attached.)

4. I had to ensure that the images had unique backgrounds. You want the subject to be the common thing between all images. Too many images with repetitive elements can bleed into the model's understanding of the character. There are images where I had to create my own background (if you see any with shitty background colours that's me.)

5. I had to organise the images so that the AI could ingest them properly.

This is just /one/ step of the process. Is this mastering the ability to draw, or learning about anatomy? No. My hands are too shaky for that. But it certainly requires a degree of skill most people don't have. This is just one step of the process. This doesn't describe the process of actually setting up Stable Diffusion, understanding the strengths and weaknesses of certain models, knowing what prompts work well for what models, knowing how to iterate on what it generates with things like inpainting, knowing what samplers work for what you're doing, and the technical skills to navigate all of this. Arguably in some regards drawing is easier. Try getting an AI to write text consistently, or draw a simple character or shape.

@w If you're doing this, you're doing w whole ton more than what the average SD or Midjourney user does, the great majority of them just use the standard models, tailoring and training your own models based on your custom needs requires skill yeah, i would say it's more of a technical skill rather than an artistic one (but yeah you have to also make a lot of qualitative selection of the images).

Of course it has it's merits, and i say it's a technical one (which i don't think it's a small thing, i know that for experience.) rather than an artistic merit because, even tough you went to these lenghts to have the model construct the kind of images you want, it's still imagery based on pre-existing data created by other artists, and there's little-to-no say on delicate and minute stuff that could make an image a work of art. Stuff like line weight, shading, stroke direction, implied meaning of a pose, everything that an artist decides to include (or not) based on the work is not there, it's based on interpolations and asimilation of existing data.

At the end my argument would boil down to: we are comparing apples to something that looks like an apple, and may taste like an apple but certainly it's not, and it's not gonna work when making an apple pie.