What is good proceedural generation?

[NotAI]

Or is proc gen dead in the market and most people, going forward, don't want anything that isn't handcrafted?

This is a spin off of the Starfield thread.

So we know what bad procedural generation is by many examples?

We also know what boring proc gen is. Plenty of it when content is made for big open world games. (It's anything incoherent?)

Then what is good procedural generation?

As in that results in play that is fun?

Dwarf Fortress? Minecraft?

Those are not rpgs. What about for rpgs?

What if there were also procedurally generated cinematics to go along with it?

Then players can bang not only bears. But also alligators, zebras, and great white sharks.

[SpellSword]

Procedural generation is a powerful tool but I'm less interested in a fully machine created product. I'm hoping that handcrafted games will continue to be valued above anything AI generated as a real person is trying to convey 'something' through it to the player.

Those are not rpgs. What about for rpgs?

[rusty_shackleford]

Then what is good procedural generation?

As in that results in play that is fun?

Dwarf Fortress? Minecraft?

Those are not rpgs. What about for rpgs?

What if there were also procedurally generated cinematics to go along with it?

Something that barely ever gets any discussion is Ultima Ratio Regum
https://en.wikipedia.org/wiki/Ultima_Ra ... ideo_game)

It's the interesting parts of dwarf fortress's procedural generation except significantly more detailed

[maidenhaver]

I don't know where the tech is at. It could have been good for space flight, or wilderness. Nothing that's supposed to be manmade. Not in a 3d game, at least.

[gerey]

I see no reason why a well-optimized LLM couldn't generate coherent narratives for side-quests in games, with their own unique twists and turns, NPCs, antagonists et al.

Stories are built on tropes, and said tropes are usually grouped together in certain combination. Also, stories follow certain patterns and beats, which an AI can also use as reference to weave a reasonably logical narrative.

Even better if you can make it take into account variables of the worldstate and player actions.

Other than hardware limitations there shouldn't be any reason an AI couldn't handle this. Even the general inability of an AI to weave a long-form narrative is not an issue since it would be only generating a few sentences outlining your objectives, with the accompanying dialogue.

Translating the output into elements in the gameworld would be far more difficult, but I don't see why any of the above would be an insurmountable technological challenge.

[rusty_shackleford]

I see no reason why a well-optimized LLM couldn't generate coherent narratives for side-quests in games, with their own unique twists and turns, NPCs, antagonists et al.

Stories are built on tropes, and said tropes are usually grouped together in certain combination. Also, stories follow certain patterns and beats, which an AI can also use as reference to weave a reasonably logical narrative.

Even better if you can make it take into account variables of the worldstate and player actions.

Other than hardware limitations there shouldn't be any reason an AI couldn't handle this. Even the general inability of an AI to weave a long-form narrative is not an issue since it would be only generating a few sentences outlining your objectives, with the accompanying dialogue.

Translating the output into elements in the gameworld would be far more difficult, but I don't see why any of the above would be an insurmountable technological challenge.

I don't think there will be much demand for AI-generated content like this outside of porn. I'm not saying the machines lack "SOVL" like paintpigs do, but they do lack intent. Playing something that was just created by a computer isn't the same as something that was purposely created with consideration put into each detail placed. You're playing a certain experience as someone, or a team of someones, wanted to design and present to the world.
I believe in the filmworld the reelpigs call it auteurship. You've probably seen(and enjoyed) this many times without putting much thought into it — I'm sure there's at least one or two Vogel fans here, after all.
There are plenty of games that are the work of one or a handful of developers, and after they leave, the studio no longer makes the same games, they just lack something that isn't there anymore. It's also a major reason for the popularity of indie games. Ironically, AI massively benefits the auteur in a way that it could never do to the mega-corporation, and is arguably a much bigger threat to them than small developers.

I suspect if I'm proven wrong it will not be because my thesis is wrong, but because games are created by AIs that have unique, intricate personalities and therefore the games they make are unique to the point where no other AI can create something like it.

[maidenhaver]

Procgen suits a game where you have a hub of dev work and a wilderness to colonize, with a radient ai assuming roles in the colonies, and a drop-in multiplayer helps. What I thought Starfield would be. I don't know where the tech is. Of recent examples, I've only played Valheim, but it can't generate coherent dungeons. If bgs had really intended to make something different from what they've done in the past, they would have built a different engine. They stagnated years ago, and will never innovate, so I don't expect advances in good use of procgen from them.

[gerey]

I don't think there will be much demand for AI-generated content like this outside of porn. I'm not saying the machines lack "SOVL" like paintpigs do, but they do lack intent.

Not denying any of those claims.

I'm not sure how much demand there will be for AI in games. All we've heard so far is a few publishers/developers experimenting with the technology. The fervor will likely die down significantly once the AI bubble bursts, but I think we're still far away from fully figuring out the potential of the technology.

And yes, LLMs are awful at generating a coherent plot, true, but with a properly tuned model you could get it to create a quest that, on the surface at least, has an interesting premise. Even something as basic as a "go there, kill critter" type of mission can be made all the more interesting if there's some sort of backstory to the target of your bounty. I'm not saying the AI should write the whole plot of the game, but merely populate it with additional side-quests, preferably ones generated based on what the player did.

We're fast approaching the limits of what any developer, regardless of time, money and manpower, can cram into a game, with modern Rockstar likely being the upper limit of what is realistically achievable or financially sound to attempt. The only two paths forward is scaling back, or making use of AI to generate content that sits in a middle-ground of sorts - not so formulaic and random to be instantly categorized as procedurally generated, and yet not nearly as good as what a talented human (or team of humans) could create.

The AI also has the distinct advantage of being able to create new content out of thin air, on the fly, and tailor made for whatever the player is currently experiencing. Also, no one has actually attempted to train a model to do what I spoke above, so there's really no proof either way on whether something like it is possible, let alone feasible.

Cyberpunk 2077 was touted as the true "next-gen" experience, and many people bought into the hype, so there's clearly an appetite for games that push the boundaries further than ever before.

[rusty_shackleford]

We're fast approaching the limits of what any developer, regardless of time, money and manpower, can cram into a game,

Strong disagree. Developers were putting more in their games 15 years ago, we've only regressed due to significant shifts in the average game developer.

---

IMO the only reason a game with the scale of RDR2 was made when it was is because rockstar was british and it merely took the mind virus a few more years to get to their shitty little island. But it shows plenty of cracks around the edges, it's like a game designed by two separate teams then stapled together.
My evidence for this is all the drama surrounding GTA6 & massive cultural changes at rockstar since.

[StudiedAsleep]

I find games procedurally generated really boring after 20 minutes.

I think it would be good as a development tool only. Could use it to generate hundreds of non-instanced player housing for an apartment building in an mmo or something.

I can’t think of any procedurally generated games that I really enjoyed.

[StudiedAsleep]

I don't think there will be much demand for AI-generated content like this outside of porn. I'm not saying the machines lack "SOVL" like paintpigs do, but they do lack intent.

Not denying any of those claims.

I'm not sure how much demand there will be for AI in games. All we've heard so far is a few publishers/developers experimenting with the technology. The fervor will likely die down significantly once the AI bubble bursts, but I think we're still far away from fully figuring out the potential of the technology.

And yes, LLMs are awful at generating a coherent plot, true, but with a properly tuned model you could get it to create a quest that, on the surface at least, has an interesting premise. Even something as basic as a "go there, kill critter" type of mission can be made all the more interesting if there's some sort of backstory to the target of your bounty. I'm not saying the AI should write the whole plot of the game, but merely populate it with additional side-quests, preferably ones generated based on what the player did.

We're fast approaching the limits of what any developer, regardless of time, money and manpower, can cram into a game, with modern Rockstar likely being the upper limit of what is realistically achievable or financially sound to attempt. The only two paths forward is scaling back, or making use of AI to generate content that sits in a middle-ground of sorts - not so formulaic and random to be instantly categorized as procedurally generated, and yet not nearly as good as what a talented human (or team of humans) could create.

The AI also has the distinct advantage of being able to create new content out of thin air, on the fly, and tailor made for whatever the player is currently experiencing. Also, no one has actually attempted to train a model to do what I spoke above, so there's really no proof either way on whether something like it is possible, let alone feasible.

Cyberpunk 2077 was touted as the true "next-gen" experience, and many people bought into the hype, so there's clearly an appetite for games that push the boundaries further than ever before.

I think there are many more features devs could add into games which wouldn’t take too much time at all, but they all just follow each other instead like sheep to play it safe and it all ends up being the same slop of shit but different story.

[rusty_shackleford]

A bunch of modders made Archolos using the gothic 2 engine and it's one of the best RPGs ever made with tons of reactivity, time sensitivity, day & night cycles, NPC schedules, etc., The issue is obviously with the people making the products, not some inherit limit in making them.

[gerey]

Developers were putting more in their games 15 years ago

Depends on the games, though.

Do forgive me for cherrypicking, but I'll be using a few examples - and I'm sure you can counter with plenty of your own, but I'm just doing it to illustrate that it can be true both ways. I'll be mostly going with average completitionist times:

System Shock 2 - 22 hours
Prey - 46 hours

GTA 3 - 38 hours
GTA: SA - 75 hours
GTA 4 - 100 hours
GTA 5 - 84 hours

Witcher 1 - 64 hours
Witcher 3 - 173 hours

Deus Ex - 39 hours
Deus Ex: MD - 52 hours

And granted, hours played isn't the best metric, since content is a fairly nebulous quantifier, and content isn't always content, but as budgets and team sizes increased, so did the amount of stuff they were able to cram into a game.

Game length is also a big selling point nowadays, since people want to feel like their purchase has value.

[rusty_shackleford]

Developers were putting more in their games 15 years ago

Depends on the games, though.

Do forgive me for cherrypicking, but I'll be using a few examples - and I'm sure you can counter with plenty of your own, but I'm just doing it to illustrate that it can be true both ways. I'll be mostly going with average completitionist times:

System Shock 2 - 22 hours
Prey - 46 hours

GTA 3 - 38 hours
GTA: SA - 75 hours
GTA 4 - 100 hours
GTA 5 - 84 hours

Witcher 1 - 64 hours
Witcher 3 - 173 hours

Deus Ex - 39 hours
Deus Ex: MD - 52 hours

And granted, hours played isn't the best metric, since content is a fairly nebulous quantifier, and content isn't always content, but as budgets and team sizes increased, so did the amount of stuff they were able to cram into a game.

Game length is also a big selling point nowadays, since people want to feel like their purchase has value.

My obvious counterargument is this is caused by the widespread adoption of the Ubisoft formula(Hello, Witcher 3.) It's a cheap, relatively easy way for designers to do paint-by-numbers open world and pad out their game to an excessive length with boring, tedious activities.
Also, GTA5 was 11 years ago, grandpa.

[J1M]

I see no reason why a well-optimized LLM couldn't generate coherent narratives for side-quests in games, with their own unique twists and turns, NPCs, antagonists et al.

Stories are built on tropes, and said tropes are usually grouped together in certain combination. Also, stories follow certain patterns and beats, which an AI can also use as reference to weave a reasonably logical narrative.

Even better if you can make it take into account variables of the worldstate and player actions.

Other than hardware limitations there shouldn't be any reason an AI couldn't handle this. Even the general inability of an AI to weave a long-form narrative is not an issue since it would be only generating a few sentences outlining your objectives, with the accompanying dialogue.

Translating the output into elements in the gameworld would be far more difficult, but I don't see why any of the above would be an insurmountable technological challenge.

I don't think there will be much demand for AI-generated content like this outside of porn. I'm not saying the machines lack "SOVL" like paintpigs do, but they do lack intent. Playing something that was just created by a computer isn't the same as something that was purposely created with consideration put into each detail placed. You're playing a certain experience as someone, or a team of someones, wanted to design and present to the world.
I believe in the filmworld the reelpigs call it auteurship. You've probably seen(and enjoyed) this many times without putting much thought into it — I'm sure there's at least one or two Vogel fans here, after all.
There are plenty of games that are the work of one or a handful of developers, and after they leave, the studio no longer makes the same games, they just lack something that isn't there anymore. It's also a major reason for the popularity of indie games. Ironically, AI massively benefits the auteur in a way that it could never do to the mega-corporation, and is arguably a much bigger threat to them than small developers.

I suspect if I'm proven wrong it will not be because my thesis is wrong, but because games are created by AIs that have unique, intricate personalities and therefore the games they make are unique to the point where no other AI can create something like it.

It also has a chance of being able to offer what games today can't: deep reactivity.

Easiest concrete example to explain would be an LLM acting as DM for a tabletop game. It would be able to take the players outside of a scripted adventure and generate encounters faster than a human. It would probably lack some overarching consistency, but be highly reactive.

Applying those same principles to a CRPG, the LLM can provide similar reactivity where a human DM is simply not an option.

In general, I think the best procedural generation is what you see in Diablo 1: hand-crafted pieces assembled by randomness.

[gerey]

It's a cheap, relatively easy way for designers to do paint-by-numbers open world and pad out their game to an excessive length with boring, tedious activities.

So why not just replace it with AI? Whatever the algorithm comes up with can't be any worse than the time-wasting Ubisoft saddles their games with.

Also, GTA5 was 11 years ago, grandpa.

Fucking Hell, really?

[maidenhaver]

I find games procedurally generated really boring after 20 minutes.

I think it would be good as a development tool only. Could use it to generate hundreds of non-instanced player housing for an apartment building in an mmo or something.

I can’t think of any procedurally generated games that I really enjoyed.

I put over 20 hrs in Valheim, which is unusual for a recent game. Granted, I didn't play with any purpose than to build a house and a boat, and die at sea, but if that came packaged with a colony builder, charters to make war on my faction's enemies, galleons to rob, and so on, you'd have a great setting for Daggerfall 2 in space.

[asf]

Strategy games have benefited from procgen since the dawn of times, from random generated maps like in civ, to side areas for resource expansion like factorio. There are other less obvious cases where things were generated procedurally initially, then the result worked upon further by hand, helping the devs as a tool. From a technical point of view, procedural stuff can be built on demand and uses less resources to maintain, which can be useful in some cases.

[1998]

Developers were putting more in their games 15 years ago

Game length is also a big selling point nowadays, since people want to feel like their purchase has value.

Yap, it used to be more about replayability, now it's about time to 100%.

[Xenich]

I see no reason why a well-optimized LLM couldn't generate coherent narratives for side-quests in games, with their own unique twists and turns, NPCs, antagonists et al.

Stories are built on tropes, and said tropes are usually grouped together in certain combination. Also, stories follow certain patterns and beats, which an AI can also use as reference to weave a reasonably logical narrative.

Even better if you can make it take into account variables of the worldstate and player actions.

Other than hardware limitations there shouldn't be any reason an AI couldn't handle this. Even the general inability of an AI to weave a long-form narrative is not an issue since it would be only generating a few sentences outlining your objectives, with the accompanying dialogue.

Translating the output into elements in the gameworld would be far more difficult, but I don't see why any of the above would be an insurmountable technological challenge.

I don't think there will be much demand for AI-generated content like this outside of porn. I'm not saying the machines lack "SOVL" like paintpigs do, but they do lack intent. Playing something that was just created by a computer isn't the same as something that was purposely created with consideration put into each detail placed. You're playing a certain experience as someone, or a team of someones, wanted to design and present to the world.
I believe in the filmworld the reelpigs call it auteurship. You've probably seen(and enjoyed) this many times without putting much thought into it — I'm sure there's at least one or two Vogel fans here, after all.
There are plenty of games that are the work of one or a handful of developers, and after they leave, the studio no longer makes the same games, they just lack something that isn't there anymore. It's also a major reason for the popularity of indie games. Ironically, AI massively benefits the auteur in a way that it could never do to the mega-corporation, and is arguably a much bigger threat to them than small developers.

I suspect if I'm proven wrong it will not be because my thesis is wrong, but because games are created by AIs that have unique, intricate personalities and therefore the games they make are unique to the point where no other AI can create something like it.

I think it is possible to use AI under direction though in terms of content creation. I mean, an AI only really knows how to create from what already exists, so... if it is trained well enough under a given direction of design, it may be able to assist in areas under that guidance. That is, how well it performs will be dependent on how much effort is put into the AI itself.

Lets say you fed an AI everything about AD&D (lore, rules, settings, mobs and their history, behaviors, past modules, etc...) and on top of that, gave it numerous examples within the constraints you are looking to produce, and established a core methodology that is to be applied, I do think an AI could create a "Dungeon" for example that would achieve a result that fits with what is expected and have a somewhat "hand crafted" feel, especially if you also attempted to define the differences in what that means to the AI.

I think at the least, you could use the AI to produce most of the content, then... the developer could go in and hand touch things to give it the corrections and feel needed. This certainly would speed up the building process of things like levels, areas, NPC creation, etc... which I think as you mentioned, would greatly reduce the workload for the small to individual development teams.

I think the people who are going to excel at content creation are those who have talent in making a good game, but also understand how to use the tools of AI to their benefit.

[Nooneatall]

We're fast approaching the limits of what any developer, regardless of time, money and manpower, can cram into a game,

Strong disagree. Developers were putting more in their games 15 years ago, we've only regressed due to significant shifts in the average game developer.

---

IMO the only reason a game with the scale of RDR2 was made when it was is because rockstar was british and it merely took the mind virus a few more years to get to their shitty little island. But it shows plenty of cracks around the edges, it's like a game designed by two separate teams then stapled together.
My evidence for this is all the drama surrounding GTA6 & massive cultural changes at rockstar since.

This is a really good point. Before the obsession with graphics devs were a lot more ambitious. Older pc games like ultima really show this off.

[Xenich]

I find games procedurally generated really boring after 20 minutes.

I think it would be good as a development tool only. Could use it to generate hundreds of non-instanced player housing for an apartment building in an mmo or something.

I can’t think of any procedurally generated games that I really enjoyed.

They had a limitation because everything had to be preplanned logic. For instance, if you were to create an infinite universe, you would have to create templates (or a list of objects, processes, etc...) and those templates would be randomly applied with some basic condition statements, the more advanced the conditions and number of templates combined with a proper seed algorithm, the better the generation was applied. That is limited though as it really is just auto-painting pre-defined content and the purpose of this was to avoid massive content creation, so spending enormous amounts of time on generational conditions would defeat the point.

With an AI, you can train it to apply development concepts, ideologies, etc... and instruct it to make decisions based on them using its basic understanding of those ideologies when it generates content. The better the AI is trained and the more knowledge it is given to pull from, the better the content will be and more "natural" it will appear compared to that of past procedural content.

The problem is, this means the developers have to put focus on the AI first, then develop the content and most who will use AI, will do it in the same manner that many apply seed generators (ie using some pre-set method already built and plug in their data which is why most random number generation over the years has been garbage).

For anyone interested and has a math background, this is a great book I read years ago on the topic (pre-AI).

Infinite Game Universe: Mathematical Techniques (Advances in Computer Graphics and Game Development) 1st Edition

https://www.amazon.com/Infinite-Game-Un ... 1584500581

[Irenaeus]

Developers were putting more in their games 15 years ago

Depends on the games, though.

Do forgive me for cherrypicking, but I'll be using a few examples - and I'm sure you can counter with plenty of your own, but I'm just doing it to illustrate that it can be true both ways. I'll be mostly going with average completitionist times:

System Shock 2 - 22 hours
Prey - 46 hours

GTA 3 - 38 hours
GTA: SA - 75 hours
GTA 4 - 100 hours
GTA 5 - 84 hours

Witcher 1 - 64 hours
Witcher 3 - 173 hours

Deus Ex - 39 hours
Deus Ex: MD - 52 hours

And granted, hours played isn't the best metric, since content is a fairly nebulous quantifier, and content isn't always content, but as budgets and team sizes increased, so did the amount of stuff they were able to cram into a game.

Game length is also a big selling point nowadays, since people want to feel like their purchase has value.

Where's the best game in the GTA series, Vice City?

[Tweed]

System Shock 2 - 22 hours
Prey - 46 hours

Both games can be completed in under eight hours and System Shock is designed that way.

I'm sure there's at least one or two Vogel fans here, after all

[StudiedAsleep]

I find games procedurally generated really boring after 20 minutes.

I think it would be good as a development tool only. Could use it to generate hundreds of non-instanced player housing for an apartment building in an mmo or something.

I can’t think of any procedurally generated games that I really enjoyed.

They had a limitation because everything had to be preplanned logic. For instance, if you were to create an infinite universe, you would have to create templates (or a list of objects, processes, etc...) and those templates would be randomly applied with some basic condition statements, the more advanced the conditions and number of templates combined with a proper seed algorithm, the better the generation was applied. That is limited though as it really is just auto-painting pre-defined content and the purpose of this was to avoid massive content creation, so spending enormous amounts of time on generational conditions would defeat the point.

With an AI, you can train it to apply development concepts, ideologies, etc... and instruct it to make decisions based on them using its basic understanding of those ideologies when it generates content. The better the AI is trained and the more knowledge it is given to pull from, the better the content will be and more "natural" it will appear compared to that of past procedural content.

The problem is, this means the developers have to put focus on the AI first, then develop the content and most who will use AI, will do it in the same manner that many apply seed generators (ie using some pre-set method already built and plug in their data which is why most random number generation over the years has been garbage).

For anyone interested and has a math background, this is a great book I read years ago on the topic (pre-AI).

Infinite Game Universe: Mathematical Techniques (Advances in Computer Graphics and Game Development) 1st Edition

https://www.amazon.com/Infinite-Game-Un ... 1584500581

There was a Star Citizen demo where they were showing off a tool they built like this for building interiors. It was pretty well done. Hopefully tools similar eventually come out for smaller studios.

[Nammu Archag]

Deep Rock Galactic has great prod gen. If you just inserted its methods of generation into an RPG you'd probably strike gold. It's relatively simple too, surprised it doesn't get mentioned or used more.

[Xenich]

I find games procedurally generated really boring after 20 minutes.

I think it would be good as a development tool only. Could use it to generate hundreds of non-instanced player housing for an apartment building in an mmo or something.

I can’t think of any procedurally generated games that I really enjoyed.

They had a limitation because everything had to be preplanned logic. For instance, if you were to create an infinite universe, you would have to create templates (or a list of objects, processes, etc...) and those templates would be randomly applied with some basic condition statements, the more advanced the conditions and number of templates combined with a proper seed algorithm, the better the generation was applied. That is limited though as it really is just auto-painting pre-defined content and the purpose of this was to avoid massive content creation, so spending enormous amounts of time on generational conditions would defeat the point.

With an AI, you can train it to apply development concepts, ideologies, etc... and instruct it to make decisions based on them using its basic understanding of those ideologies when it generates content. The better the AI is trained and the more knowledge it is given to pull from, the better the content will be and more "natural" it will appear compared to that of past procedural content.

The problem is, this means the developers have to put focus on the AI first, then develop the content and most who will use AI, will do it in the same manner that many apply seed generators (ie using some pre-set method already built and plug in their data which is why most random number generation over the years has been garbage).

For anyone interested and has a math background, this is a great book I read years ago on the topic (pre-AI).

Infinite Game Universe: Mathematical Techniques (Advances in Computer Graphics and Game Development) 1st Edition

https://www.amazon.com/Infinite-Game-Un ... 1584500581

There was a Star Citizen demo where they were showing off a tool they built like this for building interiors. It was pretty well done. Hopefully tools similar eventually come out for smaller studios.

Yeah, I see AI eventually becoming another massive shift in the industry, much like how Music production became more independent and individualized, essentially killing off the massive record companies control over things. It is already that way to an extent with ease of IDE's out there, but there is still a major gap between AAA quality and that of individuals or small studios as it concerns "production" quality and time to produce. In time, I think AI will close this gap and you will see 1 man, or very small teams developing AAA quality and scope games using AI assistance. I see the same happening with Movie production as well.

[somerandomdude]

Procgen with "seeds" is what keeps games like Battle Brothers interesting. I much prefer that to hand crafted level & world designs by talentless hacks.

[Nammu Archag]

Procgen with "seeds" is what keeps games like Battle Brothers interesting. I much prefer that to hand crafted level & world designs by talentless hacks.

I've found that GOOD procgen often feels more organic and natural than most hand-crafted worlds

[WhiteShark]

I've found that GOOD procgen often feels more organic and natural than most hand-crafted worlds

What are some examples of good procgen?