This is a very interesting use of the technology, which I hadn't thought of before.
I would really like to see the tool open-sourced. I guess that re-training the networks on a custom dataset could quickly lead to some interesting results. Even in its current incarnation, the tool could be quite entertaining.
It might also be interesting to apply the methods here to 3D environments, which are typically designed with placeholder objects and textures. I wouldn't be surprised if this were to be applied in videogames to save both on development time (art costs), and file size (multi-gigabyte texture files seems to be quite common), as a complement to procedural generation.
Given nVidia's recent publications [1] on this topic, it seems like they are not too far away from offering that kind of facilities to partner game developers.
1 comment
[ 4.4 ms ] story [ 12.8 ms ] threadI would really like to see the tool open-sourced. I guess that re-training the networks on a custom dataset could quickly lead to some interesting results. Even in its current incarnation, the tool could be quite entertaining.
It might also be interesting to apply the methods here to 3D environments, which are typically designed with placeholder objects and textures. I wouldn't be surprised if this were to be applied in videogames to save both on development time (art costs), and file size (multi-gigabyte texture files seems to be quite common), as a complement to procedural generation.
Given nVidia's recent publications [1] on this topic, it seems like they are not too far away from offering that kind of facilities to partner game developers.
Also of interest related to this software: https://nvlabs.github.io/SPADE/ https://github.com/NVlabs/SPADE
[1]: I'm mainly thinking of https://www.lyrn.ai/2018/12/26/a-style-based-generator-archi...