Finally some usages of Deep Learning that can actually enhance production speed

That’s cool.
Problem with Substance is that first you have to author textures in Substance format in order to reap the benefits of ‘procedural variations’ while that Nvidia ‘Texture Multiplier’ feature allows you to have infinite variations of pictures and even conceive new ones.

There used to be a “resynthesizer” plugin for gimp, that could generate a texture based on sample. There used to be a paper about that too. I also wrote a test doing something similar.

Yeah, but this one is more based on reconstruction based on a database of stored reality patterns instead of raw pixel processing techniques.

Resyntehsizer plugin and related paper basically used source image as a “database”. It wasn’t exactly a “raw pixel processing”. The paper related to it also discussed generation of infinitely tileable texture mappings based on non-tileable samples. It actually used neural netwrok for speeding the process up.


Checked the links. Linked PDF papers mostly talk about “Super resolution” :-\