Praguepride, it has been rules that AI generated art can't be copyrighted.
Also, be aware many artists see this as a way to get around their intellectual property rights, and is being seen by many as a way to push artists out of the equation.

https://kotaku.com/ai-art-dall...copyright-1849388060

I highly doubt any of this information is going to stop anybody, but at least try to see where artists are coming from.

So the actual training model isn't restricted against copyright but the AI just naturally has problems exactly duplicating something that fits a rigid pattern. For example it can make star trek LIKE things but you can't really tell it to make a perfect star trek insignia unless you spend the time to train it on hundreds or thousands of images to cement that look.

Many public facing models like Crayon, DALLE, Midjourney put on filters that weed out copyright or other illegal/NSFW but it's not that the model itself 'can't' do it, but you put checks in place to kick out things. Violence is a big one because nobody wants the publicity of having their AI model publish violent material either.

As for the artists, they're kind of SOL. The copyright laws are still being developed but similar issues involving training AI datasets have gone heavily towards the researchers. While the artists didn't give permission to have their work incorporated, training an AI model isn't directly competing with the artists so it typically falls under "Fair Use".

In addition if I go to a gallery and see an artist's work and then go home and draw a picture in that same style I'm not breaking the law unless I am directly lifting stuff from their work.

AI image generation does not copy work, but is inspired by it. I do agree that there are going to be tons of lawsuits and case law that has to be published and current trends could break against AI researchers providing avenues for artists to get their works out of training sets but honestly, the cat's out of the bag. Even if they stop it tomorrow there are so many open source models floating around that are already trained on these same pieces of art that it's kind of too late.

It's equivalent to the RIAA or MPAA trying to crack down on the free share of music or films with the addition that the law is fuzzier because people are trading around not exact copies but music/films "kind of like" existing material.

It's interesting to watch and as I work with AI/ML professionally I have a keen interest to stay on top of the case law.

tl;dr: So far in the past law suits have not gone in favor of people suing AI researchers so I doubt the artist community will be able to stop it either. Progress marches on etc. etc.

I just wanted people to be aware of the issues if they are going to use it.

Totally fair. In respect to the artists I specifically do not put artists "names" into my prompts even if it is legal doesn't mean it is moral.

How convoluted can the description be? What can be done with something like:

Character is a tall, well-built, hybrid Vulcan Orc, trained to play a mandolin and sing to make a living. Training further to be an assassin, he incorporates a lightsaber into his mandolin which he uses to execute his targets. Dressing as a courtier, he frequents the courts of the nobles and plies his trade in the darkness; the only light upon his kills is the glowing blade springing from his mandolin.

You can put that into a prompt and it might work out well but the  AI doesn't have a strong association of how to put the "pieces" together. So if you put "trained to play the mandolin" it will likely just key onto mandolin so you don't really need the other parts. That being said the NLP (natural language processor) is very powerful so if you put "trained to  play the mandolin" it might do a picture of someone practicing the mandolin.

Just for fun I put that entire phrase into a quickie generation and this is what came out:

https://labs.openai.com/s/BGhGPIjJz03VUJSUJY4Ymm6D

So you get:

- Vulcan (pointy ears)

- Orc

- Playing  the mandolin

And  that's about it. Maybe the other stuff had an impact but at least  from our perspective most of the other stuff fell off.

https://labs.openai.com/s/TNxNwlQBjAlSyd1TVpQ6n9EF

That is another result and you can see it added "dressed as a courtier" and  "blade" even  though it skipped lightsaber.

This is why some engines use weighting so you can have more control over how you pull it in one direction or another.

It makes sense (to me, at least) that a lot of that information would be dropped by an AI...it's describing a behavior, not an appearance (everything after 'dressing as a courtier', pretty much, and even that is questionable).  If you were describing a specific scene, perhaps, but this particular AI is generating portrait work, not capturing the characters in media res, so the words preceding the description of what he does, where he does it, and how he does it would likely be interpreted by the AI (if it's got any sort of semantic filtering) as not associated with the actual appearance of the character and therefore extraneous information that can be ignored.

If it was rephrased, perhaps, it would be more easily incorporated, but I suspect that 'dressing like a courtier' is rather vague for an AI...a courtier from where?  And when?  Of what rank?  In whose court?  All of these could massively impact what 'dressing like a courtier' means, so without specifics to clarify, the AI doesn't really devote much 'effort' to realizing that part of the description.

In reply to praguepride (msg # 33):

Thanks! That last one actually looks pretty good.

I took a quick break from doing requests to take care of some personal projects of my own and figured I would share them. I have a sci-fi crew and I generated images for all of them.

Ex-military captain: https://i.imgur.com/3d6VWo4.jpg

Femme Fatale XO: https://i.imgur.com/HoMplee.jpg

Grizzled "I'm too old for this" engineer: https://i.imgur.com/MnIdYuo.png

Tough no-nonsense orbital drop trooper: https://i.imgur.com/luW7NzD.png

Spoiled rich young VIP: https://i.imgur.com/6pbWKVr.png

When you're back to taking requests, how about this one?  You might recognize her.  :-)

extremely attractive gnome female
 sorceress
 short black hair
 skin the color of coffee with cream
 green eyes
 red dress
 carrying a wolf puppy
 night sky background

Feel free to post it here, if you like.
This message was last edited by the user at 21:24, Tue 25 Oct 2022.

In reply to praguepride (msg # 43):

Those look really cool! Well done

So I'm still baking and will look at the results but from the very first batch came this which I thought looked awesome:

https://i.imgur.com/Lt39zqk.png

Don't know if it necessarily fits but I'm happy it got the wolf pup in there without being an absolute mangle.

Awesome!  Now you're going to have to let Nezaglomie keep the wolf pup to train and wear that cool-looking armor.

Alright this was a good batch, lots of good ones to pick from (including the above one)

Holding the white wolf: https://i.imgur.com/ldaB0R8.png

Did I say Green Eyes? I meant green HIGH-lights: https://i.imgur.com/xQkXCNQ.png

She's giving you the evil eye: https://i.imgur.com/rs96U0T.png

Who's a good boy! You're a good boy: https://i.imgur.com/DZG2Mst.png

Golden is the new green: https://i.imgur.com/cVrvr1i.png

It looks good but I just imagine her getting up an hour before everyone else to carefully do her eye makeup before the rest of the adventuring party wakes up: https://i.imgur.com/HcH1Qtr.png

Fun Fails:
Wolves are hard so I just gave your girl some  dog ears: https://i.imgur.com/YzsW8Mc.png

Wow!  Thanks!  I definitely like this one the best:  Golden is the new green: https://i.imgur.com/cVrvr1i.png

Most of the others don't look enough gnome-like.  She's supposed to be only 2 ft 2, so no long neck and even a wolf puppy should seem pretty big in comparison.

Because of how over-indexed WoW is, any time I mentioned "gnome" the thing got weird. If I'm up for it I might take another crack at it by re-jiggering some things around.

Another batch cooked up last night:



One of the things I've found useful is to find a famous celebrity that kinda meets the look you're going for and dump it at the very end with minimum weights to kind of "pull" the picture together. The AI can struggle if you put the celeb too high because most of the pictures it has for celebs are "red carpet" pics that can result in some bizarre (and hilarious) compositions but if you put it at the end so the AI thinks "like this guy...but subtle" it can be useful.

I tried it with a few different celebs:

Clive Owen: https://i.imgur.com/x2R7QWC.png

Jeremy Runner: https://i.imgur.com/gYq9TzN.png

The requestor suggested Daniel Craig and I think that was the right call:

Some final version:
https://i.imgur.com/GuxZvEp.png

https://i.imgur.com/rdWF8pS.png

https://i.imgur.com/cfOsTUv.png

https://i.imgur.com/bNGVX5F.png

https://i.imgur.com/XO3UffF.png

They all have varying degrees of resemblance but I think are distinct enough that it isn't just "Daniel Craig AS..." maybe more like his his brother or cousin....

Those turned out really good! (Yeah, it's my character...)

I may have to add one of these to the character description in the game...I'm still not completely sold on Daniel Craig as an archetype, but it's going to be hard to find someone else that beats some of those!

What is the name of the App that creates these? Which is the best.

There's a ton of information (maybe too much!) in these links, both on different apps and tutorials:

https://www.reddit.com/r/StableDiffusion/wiki/index/
https://www.reddit.com/r/Stabl...sion/wiki/tutorials/

Hoooooo boy. "Which is the best" is just asking for trouble.


Right now there are 3 Big Boys in the paid "AI Art As A Service" world: Dreambooth, Midjourney, and NovelAI.

NovelAI is known mainly for doing anime styles. Midjourney is primarily known for it's more evocative artistic flair although with the right prompts it can do various styles and photorealistic pictures. Dreambooth is kind of the "steady workhorse" where it doesn't necessarily specialize in anything but it does everything pretty well and also has a lot of features to help the user feed in an image (or images) so it can, say, put you into the picture.

Then if you are technically savvy and have a beefy enough computer you can run Stable Diffusion (the main model behind it all) on your local machine or use a Google Colab account to basically borrow free CPU cycles to do your own images but with the INFINITE SETTINGS you need to tweak to do it yourself (the paid services typically take a lot of the settings out of your hands so you just have to focus on the 'prompt').

There are others: DALLE-2 is pretty cool, Craiyon is a freebie one that is fun to play with and more and more are popping up every day but DB, MJ and NA are the three big boys for paid services.

As a warning paid AI doesn't look too expensive at  first ($15 gets you like 200 images through MJ for example) however it is very addictive and you can burn through a LOT of those credits as you're tweaking and refining prompts. Dreambooth gives you like 50 credits for free as a demo, MJ gives you 25 for free as a trial and I burned through both of those in a couple of hours.

Thank you for the excellent review and guidance!

 Just as a heads up for anyone with a significant visual issue, and this is just so far, but the big three previously mentioned don't work with a screen-reader.   This probably isn't a huge surprise, but in order for you to not beat your head against the wall, and instead let me do it for you, this just FYI.

For you fully sighted folks, disregard.

Me to.

So it's been awhile and I'd like to say that both my skills and the tools available have improved dramatically. Looking back at my original works here even just a few months ago the difference is night and day.

https://i.imgur.com/nwaaV8z.png

https://i.imgur.com/q0BIyqT.png

https://i.imgur.com/y5hk5yA.png