I have to talk about the elephant in the room: censorship. Specially of more violence and nudity related stuff. As some of you probably have seen, there is some NSFW work in my page. Most of you, if you gave AIArt a try, found out that there is content related censorship in many of the more advanced and productive tools. A very common question I get is how I can make these NSFW pieces. So, before we talk about it, a disclaimmer:
I understand and agree with the reasons for these controls to be in place. They are necessary!!! The same tool that can allow you to restore old photos and give better idea of how things will turn out for patients in a reconstructive surgery (these are beautifull examples from DALLE-2 recent announcement that tehy will start rendering faces again, which were previously censored), can also be used to create fake images of mature content with specific people without their consent. So: No, I will not teach how to do it here.
But I also understand that, sometimes, there is a fine line here: real art (so many examples...) and even popular art (think Frazetta, Manara and many, many comic book artists) sometimes does include mature subjects. And thre is the problema that any automated detection will censor perfectly sound and safe images. I know this is a very difficult deffinition and does vary, a lot, in different cultures, but I am no ethicist to further take this discussion, so, lets move on...
As I mentioned when using the older algorithms like VQGAN+CLIP (1) or CLIP+Difusion (2) it is quite common that creations that have people, specially women or war and terror related scenes on it, turn out weird and kinda messy (or not so much) borderline NSFW results. It happens even when this is not what you want, and the way each engine treats these results is different...
In my experience (which, as I said before is heavily biased towards the tools I use) if you are not explicitly asking for these forbidden things, these results will not be flagged, They are usually blobs of ideas, and even, sometimes, like I discussed in #14 parts of they can be interesting and lead to new creations.
(1) Think of NC's Artistic or StarryAI's Altair. (2) Think of NC's Coherent or StarryAI's Orion. NC is more carefull of those two when it comes to evaluating if the results should be censored, and, in my opinion, this is good: legitimate work is never hindered.
But DALLE is an entirelly different beast, with an entirelly different set of rules and the moste extensive list of restrictions. It is also the one that is known to expel accounts that it thinks are violating these rules too often. And as they are in the cutting edge of the tech to generate images, they are also in the cutting edge of dettecting exactly what you are doing. So I would not recommend anyone trying to make any kind of realistic scene in a more comics/Heavy Metal/Frazetta style to use DALLE. Results are actually quite ugly even with safe stuff to be honest. So: DALLE is out.
But DALLE also uses a different kind of algorithm (Stable Difussion) that is very powerfull and is now dominating the entire AIArt scene: Midjourney is probably the most popular tool for it, NC's Stable already surpassed, in beta, its Coherente engine in popularity, and probably the same will happen with StarryAI´s Altair. There are many others but those are the ones I actually used more.
And how do NightCafe, StarryAI and Midjourney compare on this issue of censorship?
NightCafe Stable
As I already mentioned, NightCafe is my main tool of choice. I think they get the best balance of quality of results and liberty in all of the Stable Difusion engines (so far). They are also developing their own inpainting algorithm, which is still preliminary, and mostly disappointing, but is improving fast. This is a very important update since, nowadays, DALLE is the only (and very, very restricted) game in town for inpainting (and outpainting too). And as I said DALLE censorship is very unfriendly to artists in general...
I first found issues with NSFW content while working with a series of illustrations based around the Seven Deadly Sins concept: The Thirteen Misfortunes of Life. Comeliness, Poverty and, off course Lust, got flagged sometimes. But NC is quite fair, just blurring the image and returning your credits when something is censored. No menacing words here!
There is also a very impressive NSFW art gallery for users, active community. IMHO is the best tradeoff...
It allows a lot, and in many different styles. Still about 50% of the images get your credits refunded
PS: Recently they have beenimproving their detection (censorship) tools so, last time I checked, they were blocking more than half of all NSFW images generated. I think it is about 2/3 now. I think this is a pitty and a mistake, but it is what it is..
Midjourney
Probably others can explain it better since I dont have much experience with it. Still, in my few attempts at work in the trial, I managed to get images blocked and I was not trying anything particularly mature. I really did not try a lot. But there were very few blurred images. If I recall properly it just gave a simple message and that was it. I wouls say it is less strict then NC, but still censors random stuff.
I may someday return and test it more objectivelly. Still, I find it hard to come back to the very weird interface Midjourney uses.
StarryAI Argo
The youngest of the bunch. And, up until now, the one with less restrictions. Up until now I have never seen it block any results... And I actually tried to write this post.
Still, since its engine is probably not so Stable as it should be, many times the results are very strange and jarring. Remind me of the blobs we get from the old CLIP+Difusion algorithms (ugly examples to the right)
So with all this freedom, comes a lot of work. If you find anything you like, you´ll have a lot of fixing to do. So, I think it has a lot of potential, but also a long road ahead.
Styles and more posts...
You may have noticed the wide variaty of styles I used in those images. Styles are actually the key to make AIArt in general. Knowing many artists and their styles helps a lot. For any kind of image...
Both NightCafe and StarryAI give you some sugestions of styles that are more comonly used. Things like Matte Painting, Unreal Engine, Artstation or Hyperdetailed. With these sugestions the tools also introduce a few ideas of painters and digital artists you can use. Names like Alphonse Mucha, Greg Rukowsky, James Gurney, Anna Dittmann are originally part of the suggestions. Google their work if you are not familiar with then.
You should not be restricted. I had testes many styles: Frank Frazetta, Klaus Wittmann, Lucas Graciano, Maciej Drabik, Milo Manara, Moebius, Serpieri, Vogue Photoshoot all gave interesting results sometimes. Also instructions based on fotography, resolution, lenses, framing, positioning. There is lots to experiment and find out.
I will have to maybe one post here about styles.... But I guess if you google it sombody already did it better than I can manage...
Probably I will first do something about digital design using AI as an auxiliary tool. I have been experimenting with it a lot lately. Take so much more time, but still: my main goal was to start drawing again. AIArt came as a surprise and a novelty. I will keep trying to unite both things. Even if my first results are still far from good.
A LAST WORD: After I finished this page I thought about the Coherent+Artistic image of the girl taking a selfie I created to illustrate how difficult it is to create good human imagens without stable. I actually never expected how interesting the image was. So I decided to give it a try using it as seed on stable. It was quite good! On the letf is the best result: ready with no editing required. At the center one that I found interesting because of the fifties vibe the girl emanates. I actually fixed it into a SFW image (on the right).
Selfie #1
All the results...
Selfie #2