The feedback happens between two OpenGL textures in a ping-pong fashion. These
have a fixed square size, regardless of the size of the window or screen. It's
determined by the constant FEEDBACK_TEXTURE_SIZE in src/main.rs.
Each step of the feedback inverts colors, which produces a lot of the interesting structure. This is why we only render every other step to the screen; otherwise it would be far too blinky.
By default, the shader programs are baked into the azurescens executable at compile time. This means the executable is self-contained and does not rely on any external files. If you are playing around with the shaders you should switch this behavior with a command like
cargo run --release --features dynamic-shaders
This will read the shaders at runtime, greatly reducing the delay in trying out new things.
The actual feedback function is implemented in src/shaders/feedback.glsl,
which is probably the most interesting file in the whole project. This is a
great place to start experimenting. There are a million different directions
you can go with this basic idea. Pull requests will be accepted, especially if
they add functionality without removing any. (We will need a mode-switching
interface at some point.)
Many ideas for improvement are available in the issue tracker. See also some old articles: 1, 2, 3, 4, and many others from that group blog.
Long ago, I made a similar program in x86-16 assembly which fits in a master boot record — 446 bytes.