Graphics
- LOD procedural generation: generate rough large-scale statistics like orbit, size, location, color + normal map, & store this only long-term
- Then generate more and more details with a neural network client-side, on the fly
- Will probably train a specialized neural network that given high-level stats & view, generates details that look identical to the high-level picture when zoomed out.
- Can train this with sth like a GAN, with training samples produced "in reverse" by running a slow algorithm that procedurally generates detailed planets, then zoom out
- No bikes or cars, or in general land vehicles: physics simulation too computationally intensive. Also who would want cars in a world where you can have hovercars
Need some fast GI approximation that doesn't rely on anything prebaked
Target: half life 2 quality on integrated graphics
Z-buffer tricks in space, possibly quadruple-precision floats in galaxy space as ground truth?
Depends on travel mechanics
Last modified 6mo ago