DiffusionStein 3-D

[andrewj]

 

Repository here: https://gitlab.com/andwj/wolf-nn

[NightFright]

So basically this is sort of an infinite level generator, like Oblige for Quake? But I see it's not about generating playable levels.

Edited November 9, 2022 by NightFright

[dasho]

Badass, and also something that doesn't require the internet or a GPU with 10 PB of RAM.

[andrewj]

11 hours ago, NightFright said:

So basically this is sort of an infinite level generator

Not really, it's more like something that can learn to mimic an existing level, and produce infinite variations on it.

 

It was inspired by StableDiffusion, a program which can generate images of almost anything based on a text prompt.  I find it really mind-blowing what it can do.  Though my computer is too old/slow and has too little VRAM to be able to run StableDiffusion on it.

 

I think we will eventually see Doom levels (or any other game) being created using a similar process -- you could train the "AI" on some maps you really love playing and it will produce new levels in a similar style.

[Wyrmwood]

This is really cool and a little bit scary, not certain how I feel about stuff like this.

[andrewj]

This animation shows a failure case of the current method, map E6L7 of the original game.  The image on the left is the original map, the ones on the right are the creations out of pure randomness, with three different pattern sizes here: 23x23, 15x15 and 9x9.  I have tried the pattern size 31x31 (the largest supported one) and it "works" but with a strong tendency to duplicate parts of the map *exactly*.  But this animation shows that smaller sizes tend to "eat" up narrow corridors.

 

 

Also the following is a run using 31x31 pattern size on a combination of three maps, which I let run for 109 steps (which took 7.5 hours to generate), and is interesting since a certain pattern seems to take over and replacing existing stuff with copies of itself (a bit like a virus).