🧵 2. GameNGen trains in two phases: an RL-agent plays the game, and a diffusion model generates frames. The focus is on human-like gameplay data generation.