AI-generated Art

Make art for user-made content.

Moderator: Forum Moderators

Forum rules
Before posting critique in this forum, you must read the following thread:
User avatar
skeptical_troll
Posts: 498
Joined: August 31st, 2015, 11:06 pm

AI-generated Art

Post by skeptical_troll »

Hi! Some of you may already know this, but there is an online platform that creates images starting from natural languages. it is called DALL-E.

True artists might be horrified by the concept, but I think it would be a great resource for UMC creators, especially for story screens. There are 50 free usages per month, which should be plenty. You just type what you want 'it' to draw and then black magic takes care of the rest.

For example, these are two examples I got after typing 'A snowy mountain with a dangerous path and a castle, digital art'. Total effort: 2 minutes of tries :mrgreen:

I checked in the conditions and it seems that the art generated can be used for anything, including commercial purposes, so I doubt there would be license problems. Just don't pretend you've done it yourself.
Attachments
DALL·E 2022-10-05 22.01.06 - A snowy mountain with a dangerous path and a castle, digital art.png
DALL·E 2022-10-05 22.01.03 - A snowy mountain with a dangerous path and a castle, digital art.png
User avatar
egallager
Posts: 576
Joined: November 19th, 2020, 7:27 pm
Location: Concord, New Hampshire
Contact:

Re: AI-generated Art

Post by egallager »

Oh yeah Kabachuha has been doing some stuff with Stable Diffusion along these lines
User avatar
Pentarctagon
Project Manager
Posts: 5526
Joined: March 22nd, 2009, 10:50 pm
Location: Earth (occasionally)

Re: AI-generated Art

Post by Pentarctagon »

Just keep in mind that copyright of this kind of generated art is not a particularly well-established area. Quoting myself from Discord:
most CC art doesn't allow someone to alter it and then say they own the copyright to that alteration, and can then place it under an entirely different license

for example, wesnoth's art that's under a CC license is CC BY-SA: use of it requires attribution and that any derivative works can only be placed under a license that's not more restrictive than the CC BY-SA.

a person can't take a CC BY-SA artwork, alter it themselves, and then say "this is mine now, and it's under a proprietary license"

but supposedly a person can train an AI on CC BY-SA art, have it generate what they want, and they can say "this is mine now, and it's under a proprietary license"

I don't get that argument

it feels more like a convenient handwaving away of the issue since it would make it a lot more difficult to use and improve these AI tools
link, or on IRC check the #general logs from 09/29/2022.
99 little bugs in the code, 99 little bugs
take one down, patch it around
-2,147,483,648 little bugs in the code
User avatar
winkr7
Posts: 53
Joined: May 16th, 2019, 2:09 pm

I got a tomb door in 10 seconds

Post by winkr7 »

You inspired me to try a story background for a tomb door.
I took all of 5 minutes to find an AI picture site and generate.
I would be interested in seeing what others can come up with.
I didn't use DALL-E, I just clicked one that come up from AI art generate in google search
yours
winkr7
tomb door.jpg
User avatar
lhybrideur
Posts: 357
Joined: July 9th, 2019, 1:46 pm

Re: AI-generated Art

Post by lhybrideur »

I am really interested in this subject for my campaign.
I am not sure I understood the discussion on License.
Does anyone know a AI image generation tool that is compatible with Wesnoth license terms? (CC?)
For example, a tool that allows me to generate an image to put in my campaign simply by attributing the image to the tool.
User avatar
Pentarctagon
Project Manager
Posts: 5526
Joined: March 22nd, 2009, 10:50 pm
Location: Earth (occasionally)

Re: AI-generated Art

Post by Pentarctagon »

So far at least, copyright can only be assigned to a human being.
99 little bugs in the code, 99 little bugs
take one down, patch it around
-2,147,483,648 little bugs in the code
name
Posts: 569
Joined: January 6th, 2008, 3:32 am

Re: AI-generated Art

Post by name »

Pentarctagon wrote: October 5th, 2022, 10:26 pm supposedly a person can train an AI on CC BY-SA art, have it generate what they want, and they can say "this is mine now, and it's under a proprietary license"

I don't get that argument
Reasonably speaking (but not necessarily legally speaking) whether an artwork is original or not comes from how much inspiration it takes from others. If a human artist creates something equally inspired by 100 different works owned by 100 different artists, that is no longer a derivative work, that is an original work that just takes hints of inspiration from others.

Now let us say the artist takes inspiration from 10000 different works by 10000 different artists. And is heavily augmenting their creative process using machine learning tools trained on inspired by those 10000 works.
User avatar
Pentarctagon
Project Manager
Posts: 5526
Joined: March 22nd, 2009, 10:50 pm
Location: Earth (occasionally)

Re: AI-generated Art

Post by Pentarctagon »

I suppose it may eventually come down to whether or not an AI can be considered to be inspired by the art it was trained on then.
99 little bugs in the code, 99 little bugs
take one down, patch it around
-2,147,483,648 little bugs in the code
User avatar
Heindal
Posts: 1344
Joined: August 11th, 2011, 9:25 pm
Location: Germany, Karlsruhe
Contact:

Re: AI-generated Art

Post by Heindal »

As far as I know the copyright from Germany, an author / creator that combines several pieces of (even copyright protected(?)) works together and creates something totally new and innovative from it, has the copyright for this new work. But this highly depends on the definition if what he creates is outstanding work or not. So I agree that this is a grey zone in a law, at least from my point of view.

(if you look how midjourney presents its beta-users as kind of content creators for putting in some line of text is somehow strange)

I browsed and tried several of these ai image to text generators:

https://www.craiyon.com

and

https://deepai.org/machine-learning-model/text2img

can be used freely, but I'm not sure if the out-come can be used. The quality however is way lower than the result of winkr7.

@winkr7 What did you use to create this image? It looks pretty much like Midjourney, but this is a closed up beta.


Here some links you might find interesting:

https://openai.com/dall-e-2/

https://stability.ai/

https://www.midjourney.com/home/

I even have more links and models, but these above are probably the most use-able things I found.
Note that even tiktok has a AI model for that, called green-screen.
The future belongs to those, who believe in the beauty of their dreams.
Developer of: Trapped, Five Fates, Strange Legacy, Epical, UR Epic Era
Dungeonmasters of Wesnoth, Wild Peasants vs Devouring Corpses, Dwarf Dwarfson Dwarvenminer
User avatar
winkr7
Posts: 53
Joined: May 16th, 2019, 2:09 pm

Re: AI-generated Art

Post by winkr7 »

Dungeon Tomb Door
Dungeon Tomb Door
Heindal

To generate these images go to https://creator.nightcafe.studio/

here is the phrase that makes each image (you can put in a random seed too)

Dungeon Fantasy Tomb for the one above already posted.

this is Dungeon Tomb Double Door
Attachments
Dungeon Tomb Double Door
Dungeon Tomb Double Door
User avatar
Heindal
Posts: 1344
Joined: August 11th, 2011, 9:25 pm
Location: Germany, Karlsruhe
Contact:

Re: AI-generated Art

Post by Heindal »

Cool, thanks for sharing this.
The future belongs to those, who believe in the beauty of their dreams.
Developer of: Trapped, Five Fates, Strange Legacy, Epical, UR Epic Era
Dungeonmasters of Wesnoth, Wild Peasants vs Devouring Corpses, Dwarf Dwarfson Dwarvenminer
User avatar
doofus-01
Art Director
Posts: 4122
Joined: January 6th, 2008, 9:27 pm
Location: USA

Re: AI-generated Art

Post by doofus-01 »

I've tried to read up on what a stable diffuser is (there is a lot of crap out there, but I'm going with huggingface.co, better suggestions welcome), and it's mostly over my head, but my current (probably wrong) understanding is that:

1. There is a step-wise deconstruction of a training image - iterations of noise-ification and the reverse function with some selected distribution, like a Gaussian. This is down-scaling or encoding (?)
2. The set of functions from step 1 are used to form a basic representation of whatever the training image was supposed to represent. The training image could be representing many things. This encoded space is where data is sorted and associated with keywords or other encoded data or functions. It's goddam magic, but it's not that different from google search, or other technology that has been with us for decades.
3. The output of step 2, the bottom of the "U-net", is upscalied/decoded from the mathematical data space into a bitmap that makes sense to (normal) humans. That upscaling is based on the models derived from the training images in step 1.

The journey from step 1->2->3 wouldn't make sense for a single image, except maybe as a sanity check. This process only makes sense if there are a lot of training images from which to build a model that can pull the data from the "user_keyword+model rules".

So, assuming that all isn't total BS, I'd say that what winkr7 and Heindal are doing is "enhanced google image search". The AI tool gets you more targeted results from a much larger theoretical image library, but it's essentially the same. Unless you know that everything going into the training set was public domain, or some permissive copyright that doesn't require attribution, you can't upload it to the add-on server. You could use it as a reference or inspiration, but you can't get effortless, instant art.

At some point we can argue about what derivative work is, since any human-drawn image has been inspired or learned from what the artist has experienced. But by the time we need to worry about that, we have deconstructed self awareness, and life in general, to such a point that internet images are the least of our concerns.
BfW 1.12 supported, but active development only for BfW 1.13/1.14: Bad Moon Rising | Trinity | Archaic Era |
| Abandoned: Tales of the Setting Sun
GitHub link for these projects
User avatar
skeptical_troll
Posts: 498
Joined: August 31st, 2015, 11:06 pm

Re: AI-generated Art

Post by skeptical_troll »

As far as I understand, there's a fundamental difference with google search which is that the images generated are new, since these neural-network based algorithms are capable of some level of abstraction (the abstract 'concepts' form what doofus calls encoding). It is closer to 'taking inspiration' than to 'copying' and even 'assembling'. Perhaps one can get things that are extremely similar to some existing art when searching for very specific thing ("the Mona Lisa") or, I presume, when the training dataset for some specific subject was extremely limited. At least in the case of DALL-E, they took measures to specifically prevent this, see here at the section 'image regurgitation'.

One way to check could be to run a google image search with these AI-generated images, to make sure it isn't close to something existing, but I suppose that's never a 100% guarantee.
User avatar
Atreides
Posts: 1055
Joined: March 30th, 2019, 10:38 pm
Location: On the 2nd story of the centre village of Merwuerdigliebe turning the lights on and off

Re: AI-generated Art

Post by Atreides »

I tried a site called wombo.art and was surprised at the results.

It seems one could also use this to create portraits.

Here is a sample for "deep one" in the cartoon style.
Attachments
Dream_TradingCard.jpg
User avatar
winkr7
Posts: 53
Joined: May 16th, 2019, 2:09 pm

Re: AI-generated Art

Post by winkr7 »

I think the style is pretty consistent for the ones I have posted.
I took fantasy and dungeon out of the names and it is still pretty close.

dusty storeroom barrel and box
dusty storeroom barrel and box
Post Reply