AI art for games

Indeed!
People seem to think concept art is a style not a discipline. It goes way beyond creating a few different images. It often involves coming up with many, many different directions, styles and explorations. And often will have tons of choices. At this point AI can only get broad strokes, you are lucky if you get the right amount of fingers. But even as it progresses. the amount of choices a concept artist may make will make increasingly complex prompts, to the point where generating the prompts/iterations will be slower than just painting/drawing. Creating a precise drawing is faster than explaining exactly how to do that drawing.
Depending on their field, they may also have to be working within certain limitations like style, existing fictional rules or creating fictional rules. And then creating many many visuals within an style/aesthetic. Buildings, Cities, costumes props, variations on those, and depending on the size of the project, style guides and rules for other artists and production artists.

Certainly when you consider some of the greats in film like McQuarrie, Chiang, Mead, Whitlatch, etc... their work is well beyond painting a few pictures. Things like MJ and such will be a useful new tool in the concept artist's tool box, but not replacing anyone. At least not professional ones. Fivver concept artists may be in trouble.

 

Fivver artist will love it more.

 

I'm not sure about Midjourney but Dall-e 2 generated outputs are guaranteed to be unique and can be used commercially (and whatever you generate is yours). I got an email from OpenAI explaining this, I'll try to find it.

 

That's great and all, but it's a bit biased. I've seen plenty of ripped assets being sold online by stores and from vendors who very clearly tell me I'm allowed to use the asset commercially.

And sure, I can... until someone spots it and sends me a friendly letter via their lawyer.

I tend to agree that an appropriately designed AI isn't merely "copying" its inputs, but I'd be hesitant to rely on that for key art. If I get a friendly letter from The House of The Mouse stating that my background imagery looks a bit like something they own and that I ought to stop using it, that's probably pretty manageable. If the main character who's face is on all of my advertising media needs to change...

To me this kind of thing is a little exciting because it makes concept art accessible to hobbyists, low budget teams, and so on. In my experience for small teams visual concepting rarely goes past a few rounds of "mood boarding", which is definitely useful, but nowhere near as good as an actual artist.

Telling my Friendly Neighborhood Neural Network what to do also isn't as good as having an actual artist, of course, but it looks like it might make the gap smaller.

(Edit: Fixed context where I'd removed some stuff.)

 

On a technical level, if each user is assigned a seed, it won't generate the same result, it's a bit like pcg, it rely on some level of noise and randomness. So on a technical level, there is a mecanism to enforce some uniqueness.

They are known to stole original design from fanworks of their works, ie in which the only copyright claim is it "could be" in their IP. They have very aggressive lawyer but they use every intimidation tactic possible to get away, so they will find an unrelated issues to attack you there instead of addressing the legality of their act. It's a case where no matter what, legality isn't the issue, it's their ability to finesse the 8tons legal hammer to get away with everything. I wouldn't assume you did something wrong automatically, they will still find way to crush you if you get in their way.

 

Lets not miss the forest for the trees, there. My point was to highlight where I may or may not use such solutions and why. The philosophy and right or wrong of the matter is irrelevant because there's always someone out there with bigger guns.

Though that does bring up a concept for another service: a "plagarism" detector. Universities and schools already use such things. The same concept seems like it could apply well here. I get some art online and I want to do my due diligence to be confident that it wasn't ripped, copied, based on someone else's design, or just generally close enough that I might get in trouble. So I upload it to a magic AI which scours all media from forever and gives me an ordered list of the closest matches of a) the whole thing and b) any distinctive portions (logos or whatever else I might pick). I imagine it operating much like Google Image Search, but covering more types of media. From that list I can then decide whether or not it is sufficiently different.

 

One challenge is that most AIs tend lean toward hyper realism unless you are very very specific in defining a style. Which ultimately may be better communicated to an artist via a mood board. If you are going for a super detailed style, and don't have someone to concept, you presumably are outsourcing the production art, yea? You can have your outsourcers do the concepting. (most production artists hanker to be concept artists.). In a small team I can see it being used for ideation and exploration on the design/production side, but if you don't have an artist, and approach an external one with some AI art, it may be unclear (and may annoy them). You won't have a stack of images that are all in the same style, maybe one or two at best. Again.. the mood board may be much more communicative.

Not saying it's not fun to mess with, but not might be an effective solution for designing a style.

There have images generated that are clearly recognizable source, rare, but it does sometime happen. A very early image I generated using DiscoDiffusion :

Other than being hella creepy.. (my ongoing experiments are trying to create a nice anthropomorphic raccoon as a fantasy cleric), it cleary sourced a watermarked stock photo, grid and all. I am sure that will get very rare.

Though separate from large companies protecting ip, one issue that has yet to be tested is people trying to protect their MJ(and other) art, and/or claiming prior art/prompts. I don't know if you have been on the places, but people are using them at a breakneck rate. New images are appearing multiple times a second. (and many are very similar, not as a problem with the tool, just lots of people are looking for the same thing) The sheer volume is staggering. (personally, I have done thousands via just light playing around). MJ claims the person using the tool "owns" the result, but that has yet to tested. If two (or more) people use the same series of a dozen word (very likely) and come up with similar images.. what happens? No one knows yet.

As much I very, truly enjoy playing with MJ (I really do love it... sooo fun). I wouldn't risk using the results for anything yet, and least until the landscape stabilizes.

Spoiler: MJ nailed it...

MJ gave me several good raccoon clerics:

 

If general copyright (like what you may do with a music disc you've bought for example) is to be taken par example, you can count on 5-10 years for laws and courts to adapt and clarify
Enough time for whole companies to raise and fall utilizing the new possibilities.

 

Past few days made me wonder if instead of solutions like midjourney it would be possible to hijack output of human brain somehow. Basically, the brain can produce amazing imagery (especially when it is tired, wants to sleep and is overheating).

Spoiler

One issue with current approach is that modern networks seem to be incredibly wasteful when learning, and are using a different learning algorithm compared to biological brain. As a biological brain does not need ridiculous amount of hours or examples to learn.

 

Someone has already made a Unity game using midjourney - you can play it here AIに絵を描いてもらったシューティング「Shoon」 | フリーゲーム投稿サイト unityroom

More info - Nao_u on Twitter: "Midjourneyで生成した絵を使って横スクロールシューティングゲームを作ってみた https://t.co/M6HUMhzKkW" / Twitter

And here is a new tool that could allow developers to become their own voice actors - Koe Recast - Home

Next few years will be crazy.

 

I'm not sure about that. How many 'frames' does our brain process per second? Presumably, we are constantly learning with every frame, that is we are cataloguing what something looks like, sounds like, etc. for future reference. That's how we will know something is just 'off', without being able to say why.

MJ is pretty much doing what I do when making art, that is I am referencing all the art I've ever seen and thinking of how to use that information to make my new idea. It's very strange to see how AI is imitating things our brain does but also fun. I am starting to see patterns and some repetition in what MJ is generating, but I also see that from real artists. For example, when I ask MJ to do anything emotional, it seems to want to depict a lone figure, back towards the viewer and gazing into a moody landscape. Sort of Caspar David Friedrich, but also similar to a gazillion video game images.

 

Well we are going there, you already shown the video I was going to share. It's a matter of time, we already do that for speech by capturing some potentialisation of the motor neuron.

 

That's kinda not what it is about.

The learning rate of modern neural networks is too slow. When OpenAI was trained to play Starcraft, it took it 200 years of playing to produce the result. The neural networks themselves also use back propagation, meaning they're ultimately based on perceptrons. Biological brain is interconnected. It feels like we're using an incredibly inefficient approach.

 

I

I would like to find out if there is a speech to text converter that can hook into unity and then generate artworks, so players can generate new backgrounds for example just by saying things. that will be a lot of fun

 

yeah i think for true "art" you still need an artist making final judgement calls.

Copy-cats can now generate copy-paste faster, and people who can't tell the difference got more junk food to eat and that's where the big money is, so if money is the goal I don't see any downsides here.

People who already have money (or are just dirtbags) and aim for some loftier nonsense goals are going to be able to use AI as well to help in concepting. But probably their non-conformist nature will allow them to spot the patterns early and then they'll be working hard to try and invent new patterns. Most of it will suck, but eventually something new is born, then all the copy-cats swarm on it.

In the end, nothing changes, just more mouths to feed.

 

Lol. You aren't the most positive person out there, are ya? xP

 

ran out of the good coffee. Drinking the brackish crap.

 

It's a Friday in the art world, he is a realist and an observer. Ultimately, prompting the ai still require a good knowledge of art and techniques to get good results out of these ai, like specifying lens, aperture, camera model and artistic current. That will quickly separate scrub from master. Also you will still need to master some techniques to paint over the imperfections in the short terms.

And like I said, text to image aren't the only ai techniques, wait till they all converge into a power tool, mix with traditional and new non ai tools.

It's also worth noting that if you have enough vram and a decent GPU. You can still reproduce decent ai tools from dataset you can generate yourself. It's still relatively accessible. For example modder have been doing this:

Check xvasynth for example using video games dump of voice as training set.
I don't know if the quality of the sound can be good enough, but even selecting 'ai voice' in a menu to prime user about the quality is helpful, it's better than no voice. Also it imitate intonation, so training one voice per emotion and splicing it correctly in post prod would be helpful. Keep actor quality for set pieces and have ai for the chatter you can now generate massively from madlib templates to get the needed variety. Which for very small studio can make a difference, we already cut corners around.

 

Another one to keep your eye on, especially if you want something to run yourself locally:
https://mixed-news.com/en/open-source-rival-for-openais-dall-e-runs-on-your-graphics-card/

 

That isn't really true. The AIs are trained intended to generate aesthetically pleasing images. You can enter single word prompts and get amazing results and even typos make great stuff. I typed "foo" in the mj prompt to see if the bot was responding and got some amazing results. ("bar" also did great).


As was mentioned about, making a great image about making great choices, and in the case of AI art, the tool is making 99.999999% of the choices. An end user of the tool is basically one die roll the end of a millions of other die rolls.

Consider this, a user can enter a long, complex set of prompts and get a very different result than another user entering the same. Even if you enter exactly the same prompt with the same seed, there will be minor differences, which can diverge. Not to mention that different versions of the model or different tools will give radically different results. In the end, even the best prompts are going to yield random and different results. The user is really just picking from different images entire created by the software. Creating prompts isn't really a skill, it is largely random results and a set of parameters that you can learn everything you need to know about in about 20mins or so.

Getting desired results is about very tiny bit of good prompt crafting, a good amount of time rerolling based on selections, a really good source image, a huge amount of luck, and a massive, massive amount of self-delusion and a bit of confirmation bias. Someone sits down to create "magical floating castle in a storm", then re-rolls, tweaks and adjusts for an hour and then claims "That is exactly what was in my head!".... it wasn't. There billions upon billions of variations on that theme. They got something they liked, or got board or something that was close enough. That's all well and fine, but the idea that people using the tool are doing much more than selecting something that was created for them isn't a skill, it is something everyone can do to pretty much the same degree with minimal effort. Getting results you are happy with via this method, doesn't mean you wouldn't get results you would be happy with using different prompts.

---

I have actually been toying with the same idea, though more of a casual/puzzle game. I was able to generate a ton of great assets for it that require minimal clean up. I am going to try to generate more for it, and then build a game around it. Heh.. though drawing the assets I have go so far would have been faster.

 

I'm going to put a disclaimer, This isn't my opinion, but the conclusion of early many adopter like Bakz T future, I recommend the multi modal podcast where they talk about the nuances.

That said, the point isn't to get exactly what you want, but to generate interesting image. Just asking in the style of Franzetta mean you are aware of that artist, that's culture. You can direct the overall feel of whatever random image you get by knowing the right cultural knowledge to appeal.

The skill is called prompt engineering and it's the hot topic with user of that tools, some people are already hiding theirs instead of sharing the prompt that lead to some rendering. People started to notice there was more to prompting when they started to realize just adding something like 'trending on artstation' had a big influence on the style of image.

In conclusion, prompting is basically the art of querying relevant aspect into your final image. It's a collaboration with the tools, it's exactly the same an art director would have with a team of artist. The art director need to leverage his culture to direct the team. Stuff like knowing art movement, technical term like framing and composition, camera setting, mood and atmosphere, and general art language, have influence.

So technically you aren't wrong, but there is more nuance than the machine does everything.

 

"prompt engineering". People always want to believe that they are special.

I think I could say it simpler: People who have a clear intent and a vocabulary to describe it can get what they want faster.


This is true in all things always. But people always want to believe they are part of some special group, so they keep making up new terms to describe the same things that have been done for ages.
It's a problem probably confounded by specialization - people who only ever did one thing think that the techniques they use there are unique, so they become entrenched with jargon. But it's all fluff-language to describe ideas that often are not more complicated than organizing clothes in a drawer.

Obviously, a professional artist has a better art vocabulary than a non-artist. But the AI still empowers the non-artist to generate results on par with the professional just as easily.

But generate a single pleasing image is a tiny element towards creating an artistic "whole". What gives art value is the humanity expressed through it. People will grow tired of pleasing compositions after they see ten of them, on its own it amounts to little.

 

If you have to become an engineer (in prompts!), the software is at fault also since it could list the most useful art styles, camera lenses (if handled), etc. in dropdowns or other UI to make it accessible to the masses. Having to read manuals and search through tutorials to understand the right prompts is a software design issue.
But these are early days, I hope they make the prompt systems better.

 

This isn't exclusive. Unity empower non game dev in the traditional sense of making your own engine. Photography allow people to generate image at the click of a single button, making portrait painter irrelevant. Art history is full of these moment, but Art was always more about the vision, and vision need specific speciality. The same reasoning apply to the explosion of art education flooding the market with talented artists, the one that rise above the tide aren't necessarily those who can make the prettiest image, but those with unique vision that meet an audience. In fact the people who make just beautiful composition end up working for non artist commissioning their vision, IE the client. And those artist often hate "the client".

Also while the guide above focus on composition to make pretty image. The ai generally understand way more than artistic stuff, it understand a lot of abstract concept.

Exactly, prompt is how vision are expressed, your background will allow you to conjure image people wouldn't have thought off. Just like anyone can take a pretty photo, but the one that have an impact tell a story beyond pretty picture. This hasn't change.

Also prompt engineering is a global ai term, it started with gpt 3 in which if we want to do specific task it wasn't trained for, you had to figure out specific prompt to prime it. For example, start with a pattern for the ai to imitate, and suddenly it can make website, or explain complex science as Eli5, or talk like a npc with a specific personality.

Here is an example that mix speech to text, text to speech and gpt3, with prompt data to set up the personality of the npc before interaction.

Read the video description for some technical details, and eventually some comments answer by the creator.

 

It's a unity forum, here is the npc ai conversation on unity with step if how it's done:

It's a whole sub genre on YouTube

 

Exactly! The tool creates images, and anyone can use it to very similar results with minimal effort. The only difference is what people are "selling", via self-promotion or whatever. The folks talking most about the "skill of prompt crafting" are the ones trying to drive traffic to their yt channels. It's like the big youtubers doing slime, they will talk endlessly about how to do correctly, the health benefits, blah, blah... it is still just slime, but they monetizing on it. Prompt crafting (engineering is a very, very inappropriate word for it) is just overhyped right now because it is new and shiny, and public at large isn't fully aware. It is changing, refining and will get even more simple to get specific results. Like professional fidget spinners or cup stackers. It will lose its shine once a critical mass of people become of aware of and realize it very simple to do, and doesn't require much effort.

 

I actually was part of something similar. If they're using microsoft cognitive services, that makes t he app online-dependent, and the delay is unavoidable. In event of outage, npc can't speak.

If they're using something like GPT-3, then this thing loses context fairly quickly, and larger models are memory hungry.

 

This artist/prompt engineer is pretty interesting - https://www.instagram.com/kris.kashtanova/

They are using meta-human for the character, and then MJ for the environment/images. I think it shows the promise and limitations at the same time. The illustrations are amazing, consistent and professional. On the other hand, it's not hard to use either meta-human or MJ, so conceivably anybody could achieve similar results.

Kind of makes it 'not as special' at the same time the end result is very well executed!

 

Here's another promising AI tool for generating textures Poly: Generate design assets with A.I. (withpoly.com)

 

War going on in my Team, one artist hates the idea of AI...oddly enough her specialty is college is 3d modeling and at university she will be doing more UI/UX and user experience. My other artist is addicted to Midjourney aand DAll e 2. He has no problem with it, has amazing prompts, and can adjust some weirdness. Of course, his art naturally is a bit weird.

Mom/Manager is in between. I am not a 2 artist and I love the AI..me addicted too. As for in a game, I might use these for icons, class icons, for scenery, for website, etc.

Technology quickly changes and as this gets better and better, we have to just adjust with it. As they go off to University, I will have to do more of the art work, at least for prototypes. Fortunately Midjourney style is very similar to our game, so should work.

But I get the issues. Sometimes it is just too difficult to figure it all out.

 

I teach design and illustration, so this is something I have to think about. How might AI affect future career paths? Having dived in pretty deeply in both MJ and Dall*e 2, my takeaway is that this is not going to replace concept artists working on real projects. It's great for a few initial ideas, but as @zombiegorilla said, it's no substitute for working on an actual project where the team needs to see exactly this kind of option in this artistic style/universe.

The areas I think are really in danger are the low end of design/art and stock image studios. I can see a self-published author easily generating a book cover with this that would be higher quality than what they could otherwise afford, i.e. $100 or so. Also, any article that needs 'a person looking with dread at the falling stock market' used to be a matter of going through stock libraries, now that is simply the art director getting AI to do it.

One interesting area where illustration has been preferred is in ads about medical problems. There are only so many models that are willing to have their likeness associated with a specific disease/malady. Illustration solves that but so does Dalle 2. The humans Dalle 2 creates look far more convincing than 3D models and are not 'real' people either.

 

Keep in mind what we see now is just the first generation that is somewhat usable.
Imagine what will be in 10 years. I'd recommend students to keep a close eye on those technologies in every case and to learn to handle such tools at least on a basic level.

Community opinions indeed are split. Some major news site used an AI image for their cover and Twitter went wild.
It's "they take our jaaawbs" all over again.

 

In case anyone is interested, Stable Diffusion publicly released their neural network.

The odd thing about it is t hat they attempted to integrate a censor into it (which would produce a rickroll picture if triggered), but obviously people already removed that, added a GUI to it, etc.

The interesting thing about stable diffusion is that it is possible to run it locally on something like 12 gb of VRAM. It is also available for free, though licvense has restrictions:

Now, this is no midjourney, but with some coaxing I was able to produce t hings like this, AND it also can take image input.

Spoiler

 

It is worth keeping an eye on, but those things are incredibly resource intensive right now.

It is possible to end up in another AI winter, or in situation where neural networks cannot be easily locally hosted and are only available as an expensive tool. It is possible to end up in a situation with some sort of unexpected law, etc.

Frankly, I'm happy that I can run stable diffusion. It is incredibly difficult to control though, but it is usable.

It does heat up my room quite a bit though.

 

Most jobs will be replaced by technology, all you can do is adapt. My first education/job was in desktop publishing, a industry that went from hero to zero when home computers became powerful enough. Some clients where still willing to pay for the difference in quality, but not enough.

How many programmers will lose their job in the future?

 

Stable diffusion is very cool. I was having a lot of issues with getting compositions out of Dall*e that were good. Feeding a starter image is just amazing. I could never get MJ or Dall*e to really match a charcoal drawing. I uploaded a drawing to replicate and was able to get several beautiful images that matched the style of my drawing. These are all AI generated:



Did the same with a painting asking for it in the style of painters. It seems this model swipes a little more directly from the artists, the painting style is pretty recognizable. But it's very cool the color palette and 'feeling' of my original is also there.



This is all quite unsettling!

 

What is actually unsettling is to recognize what all is terribly wrong with particular images like those. Shows very well that those AI models have "understood" very well what our human brains actually focus on first and manage to make that seamless.

 

It can also replicate multiple art styles, however, dealing with this thing makes me feel like Azimov's robot psychologist.

It hallucinates based on keywords, and some keywords can be stronger than others. Meaning if you mention specific thing, it laser focuses on it and often ignores everything else.

That's kinda the point. The goal is not to produce a perfect result, but the one that can fool the humans.

 

All the models seem great at understanding what a ‘face’ is and pretty good at how an arm works. But they really struggle with hands. They are ‘hand-like’ but not aware of how many fingers we have, or even how many arms, so this weird arm-soup, but rendered very expertly.

It’s also worth noting that students struggle with drawing hands too. But at least they know how many fingers we have!

What I found surprising in stable diffusion was that it seemed to replicate the directionality of marks in a charcoal drawing. None of the other models did that at all, just a soup of marks or no marks at all. I thought that would take a while for AI to figure out, but apparently not.

 

On related note, after viewing many generated images, I've noticed that there's a lot of reuse of patterns going on. For example, if on one drawing a specific point was a tip of the nose, the NN can reuse that tip of the nose as an elbow. A wrinkle on a skirt can turn into a wrinkle on a scarf (same position, same geometry), a dent on the wall become a major crease, a neck can turn into waist and an eye of a monster blends into background to become some minor element.

It seem like neural network has a library of a million of pattern textures and it produce every drawing by blending those using transparency. Which is very different compared to how humans draw.

On other hand, to notice this you need to view many images, preferably with the same seed.

 

None.

People get scared that they loose jobs, when industrialization came. Industrialization brought more jobs.
People were were scared that robots will take their jobs. More jobs were brought in into industry.
People get scared about AI taking jobs. Tons of data scientist and NN position were created world wide.
Artist will have more tools in hand to produce their art, regardless of AI progress.
Programmers still will be writing programs. They may be use AI tools to assist on work. But AI won't be replacing any jobs.
More specialize roles are and will be created.

 

On related note, messing with Stable Diffusion makes me remember Stephen King's novel called "Ur".

Imagine having access access to a ebook with ALL literary works exist, existed or will exist in our world and in all parallel worlds.

That' the feeling.

While it is not exactly "infinity images", it ends up being overwhelming. Because there's huge number of possible variations, and on top of that huge number of parameters.

Spoiler

I uppose it means midjourney folk will be raking in money with an excavator. There's no way that their 200 images per months on basic plan would be enough for many people.

 

Something like https://lexica.art/?q=seamless+brick is kinda interesting
On one hand, it's close to what you want as an albedo map
On the other hand, the seamlessness is faulty, and you lack everything that isn't albedo (normals, height, smoothness/specu etc)

EDIT: Probably worth checking out that site in general, @ https://lexica.art/
It's a searchable set of prompts + results that people have used with the Stable Diffusion network, and some game-dev related terms give results

 

Your image and others still shock me in how dramatic and well composed they are. I've been an artist/illustrator for many years and can usually find many problems in composition, lighting, etc. in my work and in other people's work.

Some of these AI images I am looking at and I just don't see what I would change to make it better. Like your ship, it's a great composition, beautiful lighting and very dramatic, would be a great album cover. I think the only issue I have with some of these are it's not always exactly clear what the 'thing is', meaning the AI renders it expertly, but 'it' is up for debate what it's actually rendering.

That's why I think it works particularly well for alien world's and fantasy-type imagery. It's all made-up things, so we aren't really supposed to know what 'it' is.

 

Well, I asked it to generate a lovecraftian spaceship and it obliged. There's a very small number of keywords, and apparently the thing that gives the most impact is specifying the art style.

The reason why I said "overwhelming" and remembered that King's novel is because it can generate millions of those spaceships. Actually it probably can generate 4 billion variations without me changing any parameters at all, but I suspect that after few million pictures (or few hundred thousands) some sort of patterns will emerge and I'll start noticing it. That's just with random seed alone. However, on top of random seed, there are several parameters you can tweak that can vastly alter final image. So there are even more variations, and on top of that you can add keywords.

And given that this thing can run locally(!) on my computer, and generating picture takes 11 seconds (It can take longer if I play with parameters), that's mind-boggling. I can just imagine someone spending hours trying it again and again and accumulating gigabytes of images. Same thing can absolutely happen with MidJourney, where there will be definitely a lot of people burning hundreds of dollars on generation. It is like getting access to library of Babel (and immediately getting lost in the infinity of works).

One possible future problem that I see is style replication.

For example, I can do this:
"young woman, full body portrait, by John Constable"

Spoiler

Some things need fixing

Now, John Constable is dead and won't complain, but I can also do this:
"young woman, full body portrait, by Laurie Lipton"

Spoiler

Or, "young woman, portrait, by Yoshitaka Amano"

Spoiler

("full body portrait" in this case very consistently generates nudity)

Both of those artists are alive, and I can just imagine someone getting upset when a "digital pseudo artist in a box" copies their style.

Regarding possibilities, there's also img2img and this sort of thing:

Spoiler

https://www.reddit.com/r/StableDiffusion/comments/x0b0rq/img2img_fantasy_art_walkthrough_video/

where stable diffusion is used for iterative refinement of a sketch. But this takes quite a bit longer on my hardware.

On other hand, artists aren't going out of business, because in case of neural network, it is quite difficult to iterate. Basically, you can't exactly say "I want this or that added at this point", it mostly generates the whole thing from scratch every time. Img2img seems to offer iterative approach, but I've not quite gotten hang of it at the moment and it is not fast, at least on my hardware.

 

That is so wierd. I was looking at Yoshitaka Amano's work and honestly, I like the AI version almost better than anything else I saw, although I could see how it's was inspired by Amano.

Back to the original point of the thread, there is bound to be copyright issues brought up with this. Currently, you can't copyright a style, but when I see people I know, like Donato Giancola being used as a prompt it seems exploitative. He told me that over the years he stopped trying to be efficient and was just spending more and more time on each piece. Now, all of that work to develop a unique look is being used as raw material to make something similar in minutes by anyone with virtually no skills whatsoever.

I don't know Greg Rutkowski, but he can't be a happy camper seeing how his look is now all over the place. Something will have to address this in terms of copyright at some point. Right now it's just images, but what is to stop authors and filmakers from having the same issues in a couple of years?

 

Well, there were also discarded images.

Spoiler

The idea with stable diffusion seems to be producing a batch of images and then pick the one you like the most. Depending on the subject, there can be a LOT of discarded images.
----

I don't think that movies would be at risk, because computing requirements would be definitely nuts. I also think that there are far more images than there are movies in existence, so there might not be enough training material.

Regarding style, this is a strange subject. At one hand, some artist can be happy to be "immortalized". At the same time they can be unhappy that somebody else is copying their style. But a human would be able to copy their style as well, and that would not be a problem, as long as they do not plagiarize or make forgeries.

So the style problem is likely to be resolved in favor of neural networks, because a human can do the same thing.

 

I think I got nearly to midjourney level with stable diffusion.

Just take a look (not recommended to arachnophobes, though):

Spoiler

Well, I know that midjourney offfers higher detail level, responds better to natural language and does not require intricate prompts, but this is still just nuts.

Infinite images, produced locally, for free.

Nuts.

 

This is amazing, and its free!

 

It is not FREE per say.

https://midjourney.gitbook.io/docs/terms-of-service