LBALab/lba2remake

Neural networks for high quality textures generation

makseq opened this issue · 1 comments

Is your feature request related to a problem? Please describe.
Some textures look bad in the game, like ground, footpaths, water, etc.

Describe the solution you'd like
Use NN for upsampling and maybe for generation of the realistic look. Like it was for Duke Nukem face.
image

Additional context
You made a great work! I'm very impressed! Thank you, guys!

Hi @makseq, thanks for the praises.
I actually had that idea in mind at the very beginning of the project, but never got around to try it out.

I can see two potential issues that would need to be addressed for implementing this:

  • The original textured are stored in a specific format, in a texture atlas with a color palette, we're currently reading that format directly in our game engine and making use of the palette. The loading path for various type of textures in the game (static objects, characters, ground textures) is different, and so is the way the atlas work. We need to unpack all that to known formats and implement different loading code paths to load high-res versions. It's a bit more involved than it looks like.
  • I don't think we want to go for a realistic style here, one specificity of this project is that we're trying to preserve the art style of the original as much as possible. To do that with upscaling, I think we'd need to train a custom model with a carefully curated dataset coherent with the sort of style we want to achieve (roughly: colorful/candy/cartoonish/comic-like style).