Is AI the future of Indie RPGs?

[Non-Edgy Gamer]

I think it's fair to say I am a "less skilled" artist

True.

yet I can't make use of it.

You can, but you don't want to use it for what it's trained for.

It's like you're here every day, crying that a dishwasher can't cook your eggs. "Why, oh why can't they just build a dishwasher that can cook eggs? This is useless! I can't see the use case! "

I actually think the niche is shovelware. People who have zero ability to draw. Low quality games, comics, low effort stuff where turnover and churn are important. There's definitely a market there.

I can see why it initially attracted you then.

 

[The_Sloth_Sleeps]

I think it's fair to say I am a "less skilled" artist

True.

yet I can't make use of it.

You can, but you don't want to use it for what it's trained for.

It's like you're here every day, crying that a dishwasher can't cook your eggs. "Why, oh why can't they just build a dishwasher that can cook eggs? This is useless! I can't see the use case! "

I actually think the niche is shovelware. People who have zero ability to draw. Low quality games, comics, low effort stuff where turnover and churn are important. There's definitely a market there.

I can see why it initially attracted you then.

We all know It won't work to train it for a tile set. I have just pointed out the limitations yet you stubbornly won't accept that it's really not that useful for most use cases.

 

[Non-Edgy Gamer]

We all know It won't work to train it for a tile set.

It could, but it would require someone with that use case to train it.

I have just pointed out the limitations yet you stubbornly won't accept that it's really not that useful for most use cases.

"Most" use cases being that single use case lol.

And yes, you've pointed out this "limitation" for many, many pages now - in this thread and others.

 

[tritosine2k]

Doubtful it works for tileset yet because good tileset's supposed to have depth to it just stuffed into a 2D plane. So you'd start with 3D.

If 3D worked it would be enough to model 400x400m segment of a cityscape & let it arrange 4000x4000 out of that then artists can spend time on debris (fun) and unique semi-exterior (hard). Sounds more acceptable and stylish than the unreal5 demoware so far.

 

[The_Sloth_Sleeps]

We all know It won't work to train it for a tile set.

It could, but it would require someone with that use case to train it.

I have just pointed out the limitations yet you stubbornly won't accept that it's really not that useful for most use cases.

"Most" use cases being that single use case lol.

And yes, you've pointed out this "limitation" for many, many pages now - in this thread and others.

No. It would need to understand form which it clearly does not.

As I say the use cases are extremely limited because it cannot carry over knowledge. That makes it a rigid tool.

 

[RaggleFraggle]

Warlords Under Siege is using AI art and it looks really unsettling to me for some reason

 

[tritosine2k]

Warlords Under Siege is using AI art and it looks really unsettling to me for some reason

common with upsampled image it's hazy AF, especially if artifical source. All this stuff is upsampled now from what's practically close to 480p VGA res.

 

[Ravielsk]

Indeed they do.

I have yet to find any convincing use case. Sure I do see a lot of lolli and furry portraits etc and occasionally some artist goes "this could be useful but Im not sure why" but thats about it. Theres no convincing evidence. No body of work done, no concrete pipeline for how to actually achieve an outcome.

It would be far more useful to have a tool that could accomplish the basics, over a patchy command line process that attempts grandiose artwork and utterly fails.

The thing nobody is really willing to admit is that the reason why the "loli and furry" stuff looks good is because its pretty uniform in its look. Both have been more or less standardized in look for at least 10 years and so there is a huge volume of material that does not deviate drastically in terms of actual content(poses, shading, the general shape of things) so training the AI to produce decent results on those fronts is fairly "easy". Same goes for icons and game devs assets as the market with them has become so over saturated in the last few years that even I have a couple hundred of those from random dev asset bundles.

As far as actual quality application goes I can so far see this as a sort of more advanced brush tool. You provide the AI with some handmade framework of what you want and then specify with a command prompt what you expect to get, pick the best result and move on. I can see it as being useful for producing large quantities of unique textures this way from a single template or as a sort of advance "style" that you feed images made in blender and let it redraw them as pencil art for example. I can see its uses but its not some sort of digital star trek replicator that can be used to just generate your game assets for you.

 

[RaggleFraggle]

Warlords Under Siege is using AI art and it looks really unsettling to me for some reason

common with upsampled image it's hazy AF, especially if artifical source. All this stuff is upsampled now from what's practically close to 480p VGA res.

It doesn’t just look hazy. The basic geometry, perspective, the whole composition just looks wrong. There are oddities everywhere that no human being would draw, not even if the artist was a surrealist. AI art has this weird quality to it that makes it immediately recognizable once you learn to recognize the pattern. I can’t articulate it beyond “wrong.”

 

[The_Sloth_Sleeps]

Indeed they do.

I have yet to find any convincing use case. Sure I do see a lot of lolli and furry portraits etc and occasionally some artist goes "this could be useful but Im not sure why" but thats about it. Theres no convincing evidence. No body of work done, no concrete pipeline for how to actually achieve an outcome.

It would be far more useful to have a tool that could accomplish the basics, over a patchy command line process that attempts grandiose artwork and utterly fails.

The thing nobody is really willing to admit is that the reason why the "loli and furry" stuff looks good is because its pretty uniform in its look. Both have been more or less standardized in look for at least 10 years and so there is a huge volume of material that does not deviate drastically in terms of actual content(poses, shading, the general shape of things) so training the AI to produce decent results on those fronts is fairly "easy". Same goes for icons and game devs assets as the market with them has become so over saturated in the last few years that even I have a couple hundred of those from random dev asset bundles.

As far as actual quality application goes I can so far see this as a sort of more advanced brush tool. You provide the AI with some handmade framework of what you want and then specify with a command prompt what you expect to get, pick the best result and move on. I can see it as being useful for producing large quantities of unique textures this way from a single template or as a sort of advance "style" that you feed images made in blender and let it redraw them as pencil art for example. I can see its uses but its not some sort of digital star trek replicator that can be used to just generate your game assets for you.

It would be useful to have a implicit shader creator for models.

Realistically It will chew up your model when applying the style. Will not work, in it's current form anyway. It's difficult to "train" these programs to do behavioural aspects(which is what it would need), they are more number classifiers.

Probably a more procedural approach or at least a weighting option for how much to apply the styling is necessary, kind of like a Photoshop filter.

Otherwise that would reduce it to a reference tool, which IDK seems a bit passive to be useful generally speaking.

 

[tritosine2k]

...

...

It doesn’t just look hazy. The basic geometry, perspective, the whole composition just looks wrong. There are oddities everywhere that no human being would draw, not even if the artist was a surrealist. AI art has this weird quality to it that makes it immediately recognizable once you learn to recognize the pattern. I can’t articulate it beyond “wrong.”

well that's a choice, former not so much. If you use movie or even videogame prompts then supposedly it's rather correct but they are parroting those same fantasy artists left and right. I use movie and actress prompts it can get interesting (mostly insane) but nowhere even close to X-files image quality even (DVD quality).

 

[RaggleFraggle]

...

...

It doesn’t just look hazy. The basic geometry, perspective, the whole composition just looks wrong. There are oddities everywhere that no human being would draw, not even if the artist was a surrealist. AI art has this weird quality to it that makes it immediately recognizable once you learn to recognize the pattern. I can’t articulate it beyond “wrong.”

well that's a choice, former not so much. If you use movie or even videogame prompts then supposedly it's rather correct but they are parroting those same fantasy artists left and right. I use movie and actress prompts it can get interesting (mostly insane) but nowhere even close to X-files image quality even (DVD quality).

Most AI art displays this wrongness upon close inspection. Structures and landscapes and basic anatomy look like they’re floating and melting

 

[tritosine2k]

...

...

It doesn’t just look hazy. The basic geometry, perspective, the whole composition just looks wrong. There are oddities everywhere that no human being would draw, not even if the artist was a surrealist. AI art has this weird quality to it that makes it immediately recognizable once you learn to recognize the pattern. I can’t articulate it beyond “wrong.”

well that's a choice, former not so much. If you use movie or even videogame prompts then supposedly it's rather correct but they are parroting those same fantasy artists left and right. I use movie and actress prompts it can get interesting (mostly insane) but nowhere even close to X-files image quality even (DVD quality).

Most AI art displays this wrongness upon close inspection. Structures and landscapes and basic anatomy look like they’re floating and melting

The opposite of that (grounded/firm) is too much to expect from the 2D approach this is (as of yet).

Just like how DLSS can't hold a candle to MSAA 16x.

 

[RaggleFraggle]

The opposite of that (grounded/firm) is too much to expect from the 2D approach this is (as of yet).

Just like how DLSS can't hold a candle to MSAA 16x.

It's not just that. I've looked at art by surrealists like Zdzisław Beksiński and Stephen Gammell. There is a very clear difference between how they draw things and how AI draws things, even though both exhibit surreal floating and melting. I don't know enough about aesthetics to describe it formally, but in most cases I can intuitively recognize whether a picture was drawn by an intelligent agent.

You know that feeling you get from Omega Mart videos where it's pretty obvious the ads and products are being outsourced to aliens with a poor understanding of how our universe works? I get an even worse vibe from AI art.

 

[Non-Edgy Gamer]

As far as actual quality application goes I can so far see this as a sort of more advanced brush tool. You provide the AI with some handmade framework of what you want and then specify with a command prompt what you expect to get, pick the best result and move on. I can see it as being useful for producing large quantities of unique textures this way from a single template or as a sort of advance "style" that you feed images made in blender and let it redraw them as pencil art for example. I can see its uses but its not some sort of digital star trek replicator that can be used to just generate your game assets for you.

This already exists. I've posted it several times. This is also how I've advocated the AI be best used.

Artists use similar techniques for their digital drawing already. 3D pose models and photobashing have been a thing for a long time now. This is just faster.

The thing nobody is really willing to admit is that the reason why the "loli and furry" stuff looks good is because its pretty uniform in its look.

Yes and no. It's simply had more specialized training, and is using VAEs in the background. These are both trained on large datasets of, as you say, somewhat similar art, but really you can generate more than one anime art style with it, so the only "standard" is that it's all anime.

As of yet, I haven't seen anyone willing to spend $50,000 of compute custom training anything other than anime, so the idea that this is the only thing SD can do is an unproven assertion. It's simply what some of the more expensive models were trained on.

edit: The Sloth Seethes rates this negatively. Imagine my shock.

 

[The_Sloth_Sleeps]

You just cannot train by lots of samples and think you are going to get a smarter system that won't ever work. There's an unlimited amount of implicit knowledge encoded within the datasets that cannot be interpolated.

For example shadows and time of day correlation. You would need to train that aspect(assuming it's possible) and then cross train back with the images dataset.

 

[Non-Edgy Gamer]

You just cannot train by lots of samples and think you are going to get a smarter system that won't ever work.

Goalpost shifting. No one is talking about a smarter system, just better results with the current system, which we're already getting.

There's an unlimited amount of implicit knowledge encoded within the datasets that cannot be interpolated.

This is what people used to say about AI art in general. Now it's here.

For example shadows and time of day correlation.

Something that you have a hard time with in your own game design by the look of it. Can we train a smarter Sloth?

 

[The_Sloth_Sleeps]

You just cannot train by lots of samples and think you are going to get a smarter system that won't ever work.

Goalpost shifting. No one is talking about a smarter system, just better results with the current system, which we're already getting.

There's an unlimited amount of implicit knowledge encoded within the datasets that cannot be interpolated.

This is what people used to say about AI art in general. Now it's here.

I really wish you'd post examples of this amazing system you keep bringing up. It truly sounds incredible. Is there really nothing it cannot do?

 

[Non-Edgy Gamer]

I really wish you'd post examples of this amazing system you keep bringing up. It truly sounds incredible.

Already have. See the Youtube vids I've posted, or the Krita or Photoshop plugin examples in other threads.

Is there really nothing it cannot do?

Make you stop seething in every AI thread.

 

[tritosine2k]

The opposite of that (grounded/firm) is too much to expect from the 2D approach this is (as of yet).

Just like how DLSS can't hold a candle to MSAA 16x.

It's not just that. I've looked at art by surrealists like Zdzisław Beksiński and Stephen Gammell. There is a very clear difference between how they draw things and how AI draws things, even though both exhibit surreal floating and melting. I don't know enough about aesthetics to describe it formally, but in most cases I can intuitively recognize whether a picture was drawn by an intelligent agent.

You know that feeling you get from Omega Mart videos where it's pretty obvious the ads and products are being outsourced to aliens with a poor understanding of how our universe works? I get an even worse vibe from AI art.

Well it can be considered the new instagram filter, except this time it's a funhouse mirror mostly. I'm having a bit of success merging Tom Clancy's Splinter Cell and Space Odyssey btw.

 

[Kev Inkline]

I give it two years and proompters will be generating game assets to the scale of like hentai/futa production of today

https://arstechnica.com/information...odels-from-written-descriptions-thanks-to-ai/

https://cdn.arstechnica.net/wp-content/uploads/2022/11/magic3d_video.mp4?_=1

Here's already a rudimentary open source distribution like Google's dreamfusion:

https://github.com/ashawkey/stable-dreamfusion

 

[The_Sloth_Sleeps]

I'm always wary of these kinds of academic papers, as academics are such bullshitters. Keywords: curated results, funding.

 

[Non-Edgy Gamer]

 

[Tyranicon]

I am both afraid and excited for the the dark future of porn that is about to unfold.

 

[Sarathiour]

I am both afraid and excited for the the dark future of porn that is about to unfold.

lol, the "porn industry" is already flooded with shovelware games, badly animated short movies and hopw roewur ne cartoon. You're just going to see a triple amount of it, and more for even the most niche shit.