Inside Roblox's Bet on World Models
TLDR
Roblox is combining a video world model called the Super Upsampler with its game engine to generate photorealistic graphics while maintaining consistent multiplayer gameplay at scale, reaching 45 million concurrent users in August 2025. The hybrid system splits responsibilities: the game engine tracks state, physics, and rules across all players, while the Super Upsampler upsamples simple engine-rendered frames into photorealistic visuals, addressing limitations where video models alone lack persistence and world understanding while game engines alone struggle with realistic graphics. This approach allows creators on modest hardware to build visually sophisticated multiplayer games without the cost or complexity of traditional photorealism pipelines.
Why it matters
Roblox is exploring video models to create photorealistic experiences without compromising scale or multiplayer performance, with an early version of Roblox Reality expected later this year. The approach uses a video model alongside an engine that remembers the world, enforces its rules, and keeps it consistent for all players.
