Neural network renders fake urban environments for videogames

Originally published at: https://boingboing.net/2018/12/17/neural-network-renders-fake-ur.html

2 Likes

They don’t have to worry about the Chicago Bean, but will they remove the Eiffel Tower at night?

5 Likes

Hey, I saw my house in there! Maybe I should have taken the Red Pill after all.

3 Likes
2 Likes

This is… this is fucking crazy. This seems like leaps and bounds beyond what’s already been done using neural networks to generate some (fairly low-res) videos of street scenes. This kind of tech could be a real game changer for video game production. In fact, depending on how the representations are stored, this could be a whole new model for how game environments are produced.

@anon24181555
I love me some procedural generation of cityscapes, but this is very, very different from something like CityEngine. CityEngine uses a set library of surface materials and architectural grammars to generate buildings. This is automatically making generic (scrubbed of specific details like signs) versions of whatever street scenes you show it, without any pre-set notions about what objects and their surfaces should be like. Presumably, if it’s like similar neural networks, it can also create novel content as well, potentially extrapolating endless cityscapes out of a bit of footage of a real location.

2 Likes

Yes I know.

Screenshot%20(1)

Would be fantastic if they could apply this tech to generating interiors on demand for open worlds. It drives me crazy that I have enormous outdoor spaces to explore, but the buildings might as well be the labyrinth walls in a first gen Might and Magic adventure.

2 Likes

Oh another ingredient for ny “Hell” simulator. An impossible everexoanding city with 720° circles and every parkingspot being taken. Also constant red lights and 24/7 rushhour.

I do wonder if the NN has any kind of spatial sense. The 720° arent a joke if the system just keeps creating new scenes in front of you without remembering what you have seen.

Also does it understand architectural style? Or is it just ploping down buildings at random? Us american suburbs wouldnt fit a European city with medieval citycore. Or even an 3 century old Timberframe next to a skyscraper.
It could be mitigated by feeding just video from one city “style”.

This is awesome stuff but as long there arent more bits and bolts its just a prove of concept.

1 Like

This topic was automatically closed after 5 days. New replies are no longer allowed.