Hacker News Reader

Comment by msoad | Hacker News Reader

I always wondered how generative ML models can be used for creating big worlds like the one in GTA. Like every tree can be unique using a generative model. Has there been any development in this area?

View on HN

Generating worlds automatically has long been done and refined, the keyword to look for is "procedural generation". The problem is that these often end up "soulless". The way around that is usually not ML but simulating the world's history: to generate interesting terrain start with boring terrain, then simulate a couple thousand years of erosion. To generate a city start with a village, then grow it over the years. To make a forest interesting simulate trees growing and competing for light for a couple centuries. This quickly gets complex and computationally intensive, so usually you end up with human artists guiding a much simpler process instead.

This has been a thing for a long time, just not with generative ML models - most solutions use rule-based procedural generation. A large fraction of trees in games have been generated this way since the mid 2000s - see https://en.wikipedia.org/wiki/SpeedTree

You don't want every tree to be unique on a map you're going to pregenerate and ship to users. It's going to bloat the size with all the tree models.

If you're going to generate them at runtime on the other hand, you need to seed it so in a multiplayer game everyone has the same tree (in case it has a hitbox or whatever), and you don't want to risk having crappy looking trees in important locations if that's just the one where the model didn't work so well.

Procedural generation of trees has been available forever, it didn't take ML to make it work, and it's generally not been considered worthwhile

It has its applications in a different way, tool assisted level design.

Software suites like Houdini, which is something of an industry standard in VFX and gamedev can potentially leverage ML techniques. AFAIK the folks who make Houdini have talked in the past about using ML to improve their terrain heightmap modelling, improving the quality of the erosion etc...

The real value add with using procedural tools to automate creating parts of large worlds (eroded terrain, placement of foliage according to rules) is that it frees up artist time to focus on hand crafting individual locations and detail into the map.

There is an interesting talk from Ubisoft which talks about how they leverage tools for building their large worlds https://www.youtube.com/watch?v=NfizT369g60 (slides: https://ubm-twvideo01.s3.amazonaws.com/o1/vault/gdc2018/pres...)

There's a great talk about Procedurally Generating Manhattan for Spiderman that's worth watching and covers things like this: https://www.youtube.com/watch?v=4aw9uyj9MAE

If you think this is interesting, all of the Spiderman GDC talks by the various staff of Insomniac Games are S-Tier.

Actually, I'd rather see some progress in housing. As in being able to go into a building, with a different unique experience each time.

But perhaps GTA is a bad fit for this kind of innovation.

Trees are nice, but I'd care much more about distinct NPCs , preferably with some ML aided dialog.

> preferably with some ML aided dialog.

Having played AI dungeon, I don't think we're there yet.

They used ML to upscale GTA Trilogy. That didn’t go amazing.

It wasn't Rockstar

I don't think this is true - most of the examples people cite look more like they used underpaid contractors to manually recreate storefronts and billboards with shoddy clip art and plain fonts. ML-based upscaling wouldn't replace the hot dog in the background here with a completely different style of hot dog like in this example.

https://www.reddit.com/r/GTA/comments/qs22ee/this_is_unforgi...

The idea that they hacked a bunch of ML models together to do that instead of just outsourcing it doesn't hold up to muster.

Reply on news.ycombinator.com