The two primary arguments I see against AI are “trained on other people’s work” and “uses too much electricity” – with a third bonus point for “artists get paid to art”.
I agree completely on all counts – but I’m also old enough to remember when synthesizers were going to ruin music and word processors were going to destroy everyone’s ability to write…
To those points, at least one song you like has either a riff from some other artist in it, or has sampled some other song outright – with permission usually – but lawsuits between artists over this sort of thing are pretty commonplace. And every book in every language contains some sequence of words that is also in some other book in some other language.
Hell, that artist you really, really like has undoubtedly cribbed anatomy or style cues from some other artist – it’s just the way the human brain works; it’s a pattern recognition engine.
The delineation between acceptable and unacceptable would seem to be the quantity of referencing and authorization to reference. Which is probably as good as it will ever get – so I thought I might try to apply that to “AI Art”.
So, argument number one – ‘trained on other people’s art’
I spend a lot of time 3D modeling, having gotten my start on the Amiga with VideoScape and Modeler 3D (the predecessor to LightWave3D) back in, oh, 1989 or so. These days I use things like Blender into Unity and Unreal and use these modelers and engines as a sort of ‘pro’ version of Poser, (remember Poser?) which I also mess with on occasion – for my own entertainment.
I also pull my creations into SecondLife for my own use… I used to sell things I made in SL, but the customer support is exhausting – so now I just model and rig stuff for my own use.
Anyway, with the backstory out of the way, my proposal for getting around the “trained on other people’s art” problem was simply to train a background and foreground LoRA (Low-Rank Adaptation) on nothing but my own stuff and a selection of public domain (PD) content to prevent model fixation… So I created hundreds of scenes in Unreal using PD or ‘free for personal use’ (PU) assets, and then created about five hundred poses of “Faye” (a character in a story I’m writing) in various PU outfits. This was all mixed with PD/PU stock scenery and photos of people.
And yes, the use of Personal Use assets means I cannot sell anything I make with them – which is fine as this is all for personal use anyway, which I will get to shortly.
While the base model (SDXL – Illustrious) was trained on who knows what, the specifics for the image generation are purely my own creations. This mimics, to a degree, the human creative process – where the brain is trained on hundreds of millions of concepts, images, sounds, etc. that are applied to a new idea to create something new based on that idea.
To clarify – you can imagine a dirt path through a pine forest being as you have seen or physically experienced one, but you have (probably) never built one out of whole cloth. Illustrious has also never built a dirt path through a pine forest with its own digital hands, but it has “seen” millions of photos of such a thing and can be tasked with distilling that exposure down to “a painterly rendition of a well worn path through a pine forest under a blue sky with a snow capped mountain in the background”…
The above is based entirely on scenes from Unity created with PD/PU assets such as Speedtree. No actual artists were harmed in the making of the above image.
Now we move on to argument number two – ‘uses too much electricity’
This one is pretty simple to solve; I run Illustrious XL V2.0 stable locally on my M3 Ultra Mac Studio. The model needs about 32gigs of system ram and another 12gigs of vram – and my Mac Studio has 128gigs of unified memory and can generate a the above image in about a minute… All while using around 250 watts – or a third of what a high-end GPU in someone’s gaming rig pulls to no-scope headshot people in this week’s shooter…
Or, in other terms, figuring 250 watts for one minute (0.00417kWh) and an average efficiency of about 4 miles per kWh for a Tesla Model 3 means the energy used to generate the above image would power your average Tesla Model 3 for 88.2 feet…
(0.00417kWh × 4miles per kWh ≈ 0.0167miles)
So, yes – using a massive datacenter to make the AI slop that’s filling the Internet is bad – but making a few images on your local Mac Studio is no worse than watching TV…
Now for the bonus point – ‘artists get paid to art’
Yes, yes they do – and over the years I’ve provided over six figures to various artists for commissions.
The problem for me has always been getting the scene in my head down on paper, over to the artist, and them visualizing the words in a similar fashion… I’m told I’m a hell of a writer and am rather amazing with imagery, but using words to export/import a visual between brains has always been fraught with peril.
It’s far more precise to hammer out a scene in Unreal and a model in Blender, and use that as a visual aid for the artist; “I’m looking for something like this – not this exactly as I want your spin on it and I love your style – but something like this.”
I’ve in fact done this a few times – but it takes hours to create the example, and even longer when those hours are spread out over a week of 5-10 minute slices dues to work and life obligations.
So, my goal with the AI thing is to ‘one and done’ the example phase. I have a selection of environments I forsee needing for illustration purposes, and a base model of the character in various outfits from the story – both in LoRA form. So now I can just take the text I would send to the artist and stick it into Illustrious, tweak as needed, and send the artist a visual example of what I’m wanting.
It’s easier for everyone involved.
And I’m not publishing anything I make as “art” – it’s not. What I’m making is basically a very fancy chart to illustrate a hypothetical concept like one would do in a board meeting.
So, the question is; is the above still subject to the hate and vitriol of the anti-AI crowd – or have I managed to find an acceptable use for AI image generation?
Only time will tell.
Leave a Reply