Info

civitAI

I’m Kuba. I cook with 3D, video, code and AI, mostly to make architecture behave in ways it actually shouldn't.

part of d.band

www.kubakub.art
© 2023 KUBAKUB.
me@kubakub.art

All rights reserved.

GAN time

So I finally trained my first GAN. After working with diffusion models for a while, where you basically guide noise into something meaningful, I wanted to try the opposite. Two networks fighting each other.
One fakes, one calls bullshit. Somehow they both get better at it.

I took 7,804 image pairs from two of my projection mapping projects: SubLatent_v3 and Foreign_v2 and split each one into its wireframe skeleton and the AI-generated color version. Then I let the network figure out how to turn one into the other.

Year: 2026

What came out is basically a machine that learned my aesthetic. The purple and orange, the impossible surfaces, that tension between clean architecture and something organic growing through it. You feed it a wireframe of any building and it paints it in my visual language. It's not perfect, but it's mine.

69 million parameters. 7 hours on a laptop. First GAN. Definitely not the last.

The surface below is the full training progression.
100 epochs, watching the network go from black noise to something that actually looks like my work.

Hover to reveal color through the wireframe. Click to zoom into a cell.
Shift+drag to enlarge an area. Scroll to dive deeper.

  shift + drag to select
  

Desktop recommended for full interaction.

Prev Next