A new AI Image model that only uses Public Domain images in its training data

[malymin]

Jordan Mayer on Twitter:

Some were skeptical that a public domain collection could produce a model with image quality rivaling top-tier AI art systems.

Here are some comparisons of early [public diffusion] outputs vs. Flux Dev, using the same prompts.

To my eyes, PD is holding its own and looking more painterly.

made with nex3's grid generator

This model (a fork of Stable Diffusion, if I understand correctly, which is an AI image generator that can be run locally from a home computer without internet/cloud access) solves one of the big issues people have with AI image generation - the fear of plagarism/art theft - by using only public domain images in its training data.

The thing about AI imaging is that... I understand why people are afraid of it, but it does have a practical use case. Much as photography competed with painting as it first emerged, and 3D animation and vector puppet animation compete with traditional cell-painted animation, AI imaging does inherently compete with illustration and digital painting...but its most practical use case, as long as proper worker protections are implemented for employees with visual art skills, is probably going to be in the same niche as clipart and stock photos. Cheap visuals that roughly suit your needs, for when your skillset doesn't include visual arts and you can't afford to hire an artist or photographer. Potentially useful for zines, indie productions, etc.

The "increase copyright protections to stop AI from being trained on living artists' work" angle that many are gunning for are a dangerous road to go down - it will have ramifications for fanart and other transformative works, while not actually protecting professional artists from being "replaced" by glorified computer programs. (And, realistically, the prompt engineers who will be paid pennies as "unskilled labor", because the program can't actually run itself. Think of how modern movies crunch the hell out of CGI post-production teams to avoid paying set engineers, lighting engineers, etc who are unionized.) If you are an employee of a corporation like Disney, your art belongs to the corporation. Making it illegal for models to train on art they don't own doesn't save a Disney concept artist from being fired after an in-house AI model is trained on their handiwork.

(Copyright is the master's tool, and it will not dismantle the master's house. Copyright did not protect Robert Kurvitz and the rest of the original ZA/UM team from having the rights to Disco Elysium stolen out from under them; it enabled it. A corporation will almost always have an easier time suing an individual for infringement than vice-versa. The power imbalance is innate. If anything, we need more intellectual property to cease being property and enter the commons, for the health of the culture.)

Even so, I think this model, showing that a public-domain only model is not just feasible but good, honestly producing more aesthetically pleasing and less "corporate" output than standard models, should be something of a reassurance. We can have "ethical" AI imaging. The issue artists need to fear isn't an ontologically evil technology, and copyright isn't the cure. (To treat it as such is a distraction that I'm sure Disney, Nintendo, etc will gladly take advantage of if we let them.) The issue we need to worry about is labor rights, plain and simple.

Comments

[stepnix]

i've run into the question more directly in TTRPG spaces where people want to use AI to generate NPC portraits and things like that, and mentally comparing that to just googling "elf fighter male"

the subset of people who would be willing to commission an artist if they can't do the first, instead of jumping straight to the second, feels... limited

[malymin]

Yea, I feel like that's the other thing...

For better or worse, people were already "stealing" art off of google without credit or permission for these things. And I've seen artists upset when they find out their art is being used for TTRPG stuff in people's home games, or even seeing it tagged as "D&D inspo" and the like on tumblr reblogs. There isn't really a lost commission being had in these cases; the people who are willing to commission bespoke art are probably still gonna do it, and the people who aren't already weren't paying artists for comissions.

It's not fair, I can understand why artists don't like it, but it means using AI for the same purpose isn't really a lost sale. And, in any meaningful sense, the AI image is probably less meaningfully art theft than the google images search results were. Certainly, it's less directly theft.

[stepnix]

yeah
yeah