Where AI Art is heading next
Every couple of months AI Art discussion explodes with a good mix of excitement and outrage over the latest engine or update, and the questions always comes back to “what does this mean for artists?” And usually the question is expressed as one of the following:
How far will the technology go?
Will the theft of art as training data stop?
Will artists be “replaced?”
I’m a developer by training, and have seen automation trends develop for a decade and a half through my career, and truth be told, AI Art is following a trend I find very familiar from other fields. And from that perspective, I’m going to give my prediction about where AI is heading next, and what impact this technology is going to have as it develops to maturity.
Questions are answered in order of how bold the answer is, so buckle up and get ready for this roller coaster!
AI Art will keep advancing
Sometimes its fun to point out the deficiencies in current AI image generation capabilities, whether it was DALL-E’s bizarre images earlier this year, or failures to generate hands or faces in more recent iterations. And it creates a false sense of safety that AI will “never really get there.” I think its a mistake to underestimate the coming capabilities of the technology.
AI developers see such shrotcomings more as a technical problem to solve than a “it can’t work,” and month-to-month the capabilities are improving. One thing I wonder about especially is “character persistence,” and NovelAI seems to enable one to keep images close enough to the original that you can, in fact, generate consistent-ish character designs from an original image.
Concerning the question of “how far AI art will go,” I think the main technical question will be wheher an AI can be trained to understand objects in 3 dimensions, from a 2-dimensional reference image. The test I’ve been giving various AIs lately is to ask the AI to take my character Pauline Censor and turn her around.
Something you might notice in this experiment is that the AI does not understand the tape. It merely reproduces it as is. Keep this in mind - this is something about AI that plays into the big question of “what does this mean for artists?”
The corporations will write the regulations
Since AI art can facilitate a certain kind of art theft, I’ve heard the calls for “regulation” already. I do think its coming, but not in the way some advocates are expecting. The regulations wil come, but they are going to come from the corporations currently getting established in the AI image space. And I can predict this with confidence because its a common historical trend.
When a new technology comes out, it always starts with a “wild west” period where there are few or no rules, and a few parties get away with something egregious. In the web development world, it was the internet security heyday in the 90s security was an afterthought and it was almost acceptable to leave a calling card on machines you hacked. A few hackers even built respectable security businesses by “marketing” this way, something unthinkable today.
In AI image generation, the most egregious was the attempted ripoff of Kim Jung Gi’s art style just days after his death. These are the moments that usually make the history books, but the real history is happening in companies building up around the technology.
In Bringing Open Source to AI Art, I sugested that one of the steps forward would be to make training data modular, including modules of training data for specific artists. Now that many art image sources are starting to close themselves to AI Bots, I think there is going to be value in offering artists money each time an AI user writes “in the style of” that artist. It would also give certain AI image companies a competitive advantage because they could offer art styles not available everywhere.
Once the first company offers artists the ability to make money by licensing work to AI, I believe we’ll be at a tipping point. This is also when companies will likewise find it advantageous for their competition to *not* rip off artists, as they would be taking “for free” what the corporation must pay for.
While there’s a lot of fear and excitement in the “wild west” period, it does come to an end when people in the space recognize their value, demand it, and companies likewise find an advantage in respecting that value. Bad actors never fully go away, but I do see a tipping point when AI image generation becomes more mainstream, and that point will be defined by AI companies increasing their value to artists.
One thing artists do that AI doesn’t
AI image generation will improve and become mroe mainstream, but this still leaves the question, will artists be reduced to generating training data for the machines? I believe the answer will be “sometimes, but not entirely.”
Working with AI image generators on experiments like the earlier one with my character Pauline has taught me one thing:
Fundamentally, “AI Art” makes derivative images from existing images via the use of statistics.
When you start making prompts at the edges of that set, or outside of it, the shortcomings become clear very fast. But so many creative works require you to do just that - go to the edges and push the limits.
AI cannot do what has not been done before.
What this means for artists is that the market for very common images, such as portaits of generic anime girls, is going to drop precipitously. However, unique character designs, concept art requiring a “signature” design language, or ensuring an image conveys a certain emotion, are all cases of things not likely to be within the scope of AI. I don’t mean “AI cannot improve image quality.” I mean that such things are outside of the scope of what an AI imgage generator is.
Like I demonstrated with the experiment above, AI does have capacity to stat-match various images and generate variations, which gives the appearance of new image. What it does not have is understanding of what it is making. And the value for an artist, therefore, will be to bring that understanding.
Therefore, I don’t think AI will “replace artists,” but I do think that it will change the artwork process by adding the ability to generate generic or derivative assets. It also is true that illustrators who rely on producing generic or derivative work will be in trouble. As a developer who deals in automation, I’ve already seen this dichotomy take shape in multiple industries, and the truth is, those who can understand the tools for what they really are, will find opportunity.
These are my predictions, what are yours? Let me know in the comments below!