What AI art can't do. And why maybe that doesn't really matter :-(

As we argue about AI (AI is evil, AI is the future, etc.) here is a thought about what "AI" actually cannot do for RPG art. And then, sadly, a thought on why that may not matter, and the potential of human-created art is going to get overlooked anyway.

This follows a walk-through of how we designed one of the first characters for The Serpent, and shows why we get better character design from humans than from computers.

The character (sketches and final) looks like this:



But as a note, I'm saying "from computers", not "from AI". Because AI is just a hype term - these things are not, actually Intelligent. And that's the point. Humans can apply real intelligence to create things.

Background: Where are we? Who is this?​

Our NPCs live their lives in a sunken domed city, surrounded by towering stone cliffs, and above the domes that seal the top of these pits a toxic atmosphere swirls. And because the feel that we want is the darker sides of classic Persian literature, they are going to be ruled over by a narcissistic monarch - perhaps charming, definitely cruel. So who is our monarch?

Draft 1: Initial sketch​

The artist (Dian) comes up with sketch, for shape, feeling, etc. And immediately we start to have a dialogue, a discussion. The archaic "robed monarch" avoids sci-fi cliche, but it feels too much like a fantasy character... it's a great idea that if a civilisation lives entirely within domes, rugged boots are pointless and comfy sandals may be preferred (even for a noble)... the crown is a bit obviously medieval so we should lose that, but the big shoulder ornament is interesting and we can build on that idea... what if we push the idea of heavy, blocky design elements.... Okay so if we want archaic Persia without being stock fantasy, what if we look specifically at... let's look at some options... oh! Parthian visuals look distinctive but archaic! And for that heavy, blocky look, what if we start looking at Brutalist architecture...?

Draft 2: Second Sketch​

We're getting a sense of who these people are, and the draft prompts more ideas and discussions. The people here will tend towards curly dark hair (from the Parthian sources, for example)... sandals for the wealthy, but what about workers - maybe flimsy footwear is a mark of status and high fashion? Clothing should be light fabrics but heavily ornamented to show status.... But then the heavy shoulder garb really stands out - but this embodies the heavy, brutalist aesthetic and hints at (with an almost Dwarven feel) the heavy metals on which the settlement's wealth is built. The prominent ring is interesting - what is it for? What if everyone here wears a ring that conveys status outwardly (material and design) but also subtly (like a digital ID card, but worn as a ring - which also provides digital and physical access)? A single ring isn't enough, we need the ruler to have more ornamentation - but now we've established that jewellery is also functional, let's wonder what, for example, an earring might actually do, functionally...?

Final Version​

Now we have a clearer sense of the culture, we have some distinctive visual traits (like the curly hair) and we're building a sense of a unique fashion, jewellery, etc. We can use this. And it adds to the richness of the world - we can even cost out and write up some of this as equipment (like the Whispering Earring) that the PCs might want to acquire.

And this journey is not one that we can take if we use a computer to generate the art.

You Can't Prompt This​

If you go to Midjourney, and prompt it to create a sci-fi Parthian king (for example) then it will give you a generic space-fantasy character - not only will it be unable to engage with the key references, like Parthian dress or brutalist architecture, it won't even be able to reference these traditions well. Not only can it not creatively interpret, it can barely cosplay. And prompting art won't help you to refine the world or add details for the players.

More knowledgeable "AI" fans will say that "prompting" is so 2025, and what we should be doing is creating context files and world bibles to train the computer. OK, fine. But in the real world, you don't create something great just because someone issues instructions and minions mindlessly execute.

Everyone reading this will have found (at work, school, etc.) that instructions given are inadequate, that the doers bring their own insights and add to the instructions, that teams get better results when they can think, reflect, improve on briefs and deliver "this and more" rather than just do what they're told. Solutions emerge as the work evolves.

And those who are genuinely creative will know that creation and discovery are interlinked. As you create, you discover and refine.

And ultimately what we've described above is a process of communication, where two (or more) people iteratively explore ideas. (There is a reason why the ad industry, at its peak, relied on creative pairings.) The idea that the "definer" has some god-like genius that can't be improved by the person who gets into the execution, is simply hubris. Without this dialogue, we get more superficial art.

But Is This A Losing Battle?​

So let's say we do this for all our art (and we will). So for every character there is a couple of hours of discussion, an iterative process of discovery... we spend half a day of time plus the time needed to actually sketch and then draw the character. And we end up with a simple, black and white ink drawing.

Meanwhile, someone else tells a computer to spawn some full colour space-fantasy images. This takes a few minutes. It is way quicker. It is way cheaper. And in the end people buying games (e.g. looking at imagery on a Kickstarter page) will think "yeah, that looks cool!" when they see the superficial colour image, and will think "that doesn't look exciting" when they see the black and white image. Which means, so long as they don't think (or care) that an image is computer-generated, they are more likely to pledge money to the "creator" who just spams computer-created content. So that "AIer" spends less on art that makes less money, and the gamer gets more superficial computer-generated books.

The economic incentives are simply against paying real people to create something nuanced, and in favour of getting computers to create something superficial. I can enthuse about the advantages of genuine human collaboration, but in the end, can we defeat the economics of "AI"?

There are hidden costs to AI that I thing business overlook. There are the large data centers with expensive hardware and power draw. The cost of internally hosting there data for said IP unless they don’t care that there training data will be shared. Even if that data is hosted in the cloud that separate infrastructure costs a lot. What happens when this bubble bursts there are too many unknowns with all of this. I’m not sure the savings from AI are viable they just hiring an artist.

HippyCraig said:

There are hidden costs to AI that I thing business overlook. There are the large data centers with expensive hardware and power draw. The cost of internally hosting there data for said IP unless they don’t care that there training data will be shared. Even if that data is hosted in the cloud that separate infrastructure costs a lot. What happens when this bubble bursts there are too many unknowns with all of this. I’m not sure the savings from AI are viable they just hiring an artist.

Click to expand...

At present, most AI is not properly monetized. In other words, people aren't supporting the full cost of the AI data centers. The costs are being paid mostly by investor money, not user payments. be real interesting to see how things shake out over the next several years...
Cold Fusion (youtuber) has a pretty good look at the current state of the AI industry.

aramis erak said:

At present, most AI is not properly monetized. In other words, people aren't supporting the full cost of the AI data centers. The costs are being paid mostly by investor money, not user payments. be real interesting to see how things shake out over the next several years...
Cold Fusion (youtuber) has a pretty good look at the current state of the AI industry.

Click to expand...

Basically, it seems they are hoping for one of three things:
1) They develop the holy grail in AI research, some actual sapient (super) intelligence. I wouldn't count on that happening soon.
2) They develop AI for specific use cases where it just beats out the traditional way and be more cost-efficient/profitable. That is likely to happen in various areas - and some uses cases it is emplyoed for right now might actually be that, if they were to raise prices.
3) They make people so dependent on their AI that they will swallow massive price increases that make things profitable. Maybe that is just 2, just phrased more negatively? Same as streaming and cloud services...

AI serves well for plugging the holes and for prototyping. It is - at this moment - incapable of serving a complete and polished product (note that there are agentic workflows that - seemingly - work well enough for some tasks to rectify that). And it's decent for managing research in large datasets (always subject to verification).

It's a terrible tool in the hands of dilettantes. Period.

The risks involved include but are not limited to, are: AI slop replacing quality (cheaper and faster), hallucinations (plausible falsehoods), intentional disinformation (hostile agencies at work), hidden human resources costs (fewer people in junior positions, for example).

Disclaimer: I use AI on daily basis, but consider it just one of many tools.

Enjoy playing with AI while it lasts. The cheap, subsidised services won't last forever. It's an interesting technology, but the current price point is only temporary. The economics just don't work.

I think that from an illustrator perspective generative AI is not iterative enough (Sketches => final design), but generative AI was not made for illustrators, as there are too few to matter from an earnings perspective, and to stubborn to get a big adoption. As the AI generative market matures that will change. Meanwhile they targeted everyone, those that didn't have an Illustrator's background and training. A far larger market, that tend to be far less critical, and more adoptive if they can do fun stuff themselves. Just like a Kia isn't designed with Max Verstappen in mind...

Different generative AI and LLM solutions have different capabilities. With ChatGPT I can toss in an image and it will work with that, reproducing that in another style, etc. But with Midjourney I can't do that (yet?), I can use collections of existing images to define a style and can work with other images generated by Midjourney. But something like Midjourney is very focused on relatively quick 'upgrades' to their models, they've had 10 versions in 4 years, It seems this has slowed down somewhat to about 1 model per year. I suspect that eventually they'll hit a development wall or they hit a threshold where it's 'good enough', just like most things related to Tech. I expect then to capability no longer be focused on depth, but in breadth, we we could see workflows that are aimed at more traditional illustration workflows. Already quite a few folks hack together flows for specific solutions to go around generative AI limits. Glibatree for example does some interesting things already...

As someone already mentioned, the current cost for 'AI' is heavily subsidized by investors, that won't go on forever. But meanwhile models get better, more efficient, hardware gets better and sometimes more efficient. On a local machine that's taking up less then a liter of space, <10W while typing this and ~70W while inferring, which is less then the old 100W light bulbs. I can run relatively large LLMs (Mac Mini M4 Pro 20c GPU, 64GB memory) on there, for example the 70b DS model gave me similar results to the half a year earlier ChatGPT 3.5. The next generation of Apple silicon (M5) is already out and significantly better then the previous generation (which was already very good). The question is will the investor money be cut off before hardware and software meet at the 'good enough' threshold? That is imho a guess for everyone, not much better then reading tea leaves.

Already companies are running into 'AI' issues, regarding security, compliance, costs, skills, etc. I'm already pretty annoyed with Apple support where every time the 'AI' agent asks what you're calling for does not recognize their own product, I have to go through the same loop a couple of times before I finally get a human on the line that, most often, can help me. I'm currently very happy with my Apple ecosystem, so I'll life with the annoying Apple AI. But I've left services before due to their support not being up to my expectations, and moved to companies that asked significantly more for their services, but their support exceeded my expectations (ISPs). Similar things will happen with other 'AI' products eventually.

I also see that if you want quality results out of xyz (even generative AI or LLM), you need someone with the right skills, and if my experience in IT is any indication, those will be pretty rare and thus expensive. Expensive AI + expensive people = very expensive! Suddenly it becomes more interesting to hire cheaper people that might take a bit longer to complete a job, but are overal cheaper. Another issue many companies are now forgetting is that they're replacing 'junior' positions with 'AI' and only retaining 'senior' positions to oversee the work/results. The thing they forget is that to become a 'senior' you need to start somewhere as a 'junior', and when no one has 'juniors' anymore because they are replaced with 'AI', suddenly when the 'seniors' dry up due to retirement or death, you have a HUGE problem. Just like with all the previous automation waves, some companies will collapse and the how the work market works will be rewritten.

That said, I'm already of the opinion that many of the illustrators in mainstream media are not that different from the generative AI illustrations. Heck, some of the generative AI output I think is better then what some of the D&D WotC illustrators produce (bland). Others like for example Paul Bonner produce very unique pieces which I love.

As for kickstarters and concept art, that's mostly an illustrators perspective, not a consumers perspective. Don't get me wrong, I've got quite a few books with artwork for computer games and tabletop games, but not everyone is obsessed with that. Let me give the example from 13+ years ago, Planetary Annihilation on Kickstarter: They had an awesome (human made) 'gameplay visualization' video that absolutely sold the concept of the game, but in the end it was NOT representative of the game, the 'gameplay visualization' was still more awesome! As a consumer I was still disappointed... Would I have cared if they made an intro movie that was equally divergent from the end product, not really. At best I would have given the generative AI the blame of not being what was shown initially. And imho that is the problem with any concept art. Not representative of the end product.

I actually like the first draft look of the art and black and white art like what 1e had. The sketch art of 3e gave things a good feel for me over a lot of other art.

That said, I do not know a lot of the art and layout design that makes one piece better or more appropriate than another piece for a specific place. It seems that a human can make a sketch and another human can choose or veto it for the place they were hoping it would land. Can I tell the AI to make something and would it be what the humans want and the same fit for the layout? Is this coming in the near future?

As an ordinary person looking at a book, how much is the art making me want the book? There is some to this and I'm guessing it depends on the person, but how much? There has been a lot of threads with many saying that they would never spend money on something but is that the same for most people? At what point is the AI art piece just as good or able to change with prompts that it would be as good to place in the book.