What's this about?
A "generative world renderer" that takes structured world states exported from a physics engine and synthesizes the RGB frames — the engine handles what happens, the neural net handles what it looks like.
What exactly is the split here?
Unlike text-to-video (generate everything from a prompt), this keeps deterministic physics and replaces only the graphics pipeline with a generative model — controllability of an engine, visuals of a diffusion model.
Why is it worth noting?
65↑ — a plausible near-term path for neural graphics in real games, sidestepping world-model consistency problems.