GAN process tutorial

MEGA EZ, EVERYTHING IS PRE-SETUP
- Ever wanted to make some of those confusing AI images where everything looks like something, but at a closer look, it's just a weird mess?
- Feel like Artbreeder doesn't give you enough control over the end result??
- How about even less control???
-
look no further
-
- You will need one of these Google Colab notebooks. Open it and copy it to your Google Drive. There you will be able to run Python code (written by other people, don't worry if you know nothing about Python) using a remote machine provided by Google. This means your local hardware doesn't matter. A free version has some limitations depending on how busy it is, but can be enough.
https://colab.research.google.com/drive/1L8oL-vLJXVcRzCFbPwOoMkPKJ8-aYdPN#scrollTo=g7EDme5RYCrt weirder results, but can render higher resolutions (I use 980 x 500)
https://colab.research.google.com/drive/1gFn9u3oPOgsNzJWEFmdK-N9h_y65b8fj#scrollTo=g7EDme5RYCrt better results, but runs out of memory with higher resolutions. (I'm using 540 x 300)
(All of the credit goes to the people who made these super-cool notebooks btw - Katherine Crowson et al mentioned in the top sections of the notebooks. We'd probably still be living in mud-huts without this kind of people <3)
-
- Build out a library of images for future use as textures or individual elements.
- Use these as rough initial sketches. Taking an entire image and doing a puppet warp pass on it in photoshop to sculpt it into something more defined works great as the first step, supplemented by photobashing.
- Do quick exploration riffing off reference images. Let's say you got a clear reference or a whitebox to base the concept from. Use that as an init image to generate endless variations. Works great with lower step size settings (more intermediate steps).
- Simply enjoy looking at the weird s**t that comes out of it. Learn to speak the machine language. Welcome our new overlords.
-
-
The images are unedited, aside from a small resolution bump.
-
Enjoy!