Roblox is exploring how a video world model can enhance its game engine to create photorealistic experiences without sacrificing scale or multiplayer performance. This hybrid approach, discussed with Anupam Singh, senior vice president of engineering at Roblox, offers insights for engineers building real-time systems.

A game's core lies in its unseen systems that maintain world consistency and render experiences. Consistency involves persistent state, fair rules, believable physics, and real-time synchronization for thousands of players, ensuring everyone shares a single, coherent world. Realism, on the other hand, focuses on fine details like moving grass or shifting light, which players increasingly expect.

While AI world models excel at generating photorealistic visuals and game engines ensure world consistency, neither is sufficient alone. A pure video model lacks durable memory, consistent rules, and reliable input-to-outcome translation, making it difficult to test and refine. It operates on pixels without true understanding of objects or their behavior. Roblox's strategy combines these technologies to divide work, maintain consistency, and achieve photorealism affordably. The platform, which reached a peak of 45 million concurrent users in August 2025, aims to make photorealistic games cheaper to build.