Roblox is building a hybrid architecture called Roblox Reality that combines a traditional game engine with a video world model (the Super Upsampler) to deliver photorealistic multiplayer games at scale. The game engine handles authoritative world state, physics, and rules, while the AI model upsamples plain engine-rendered frames into photorealistic output. The post covers four unsolved engineering challenges: latency (targeting ~30ms frame generation via self-forcing autoregressive generation), long-session consistency (extended context windows), multiplayer synchronization (server-authoritative state shared across per-player model instances), and creator control (a game cartridge harness wrapping deterministic logic around the model). Roblox acquired Morpheus AI and Lucid AI to accelerate these efforts. An early version is expected later in 2025 or early 2026, with the goal of enabling small studios to produce large-studio-quality visuals on ordinary hardware.
