PixVerse launches R2 real-time world model with persistent session memory
- PixVerse released R2, an upgrade to its real-time world model that generates a world which keeps running while a user is in it instead of returning a fixed clip like typical AI video.
- R2 splits its architecture into Omni Causal AR, which learns across more data, input types, tasks and longer sessions, and a real-time acceleration layer that compresses the same model so capability gains do not slow it down.
- In PixVerse internal evaluations, R2 cut visual drift over a long session by 35.8%, and PixVerse says it keeps state far longer than R1, the first real-time world model it shipped in January 2026.
- Inputs in R2 persist within a session: a player moves a character with WASD and arrow keys while live prompts add elements or alter the environment, and earlier actions stay true later in the session.
- Creator Xiaolongbao built Zero Mark, an interactive film-game on R2 where offering a dragon or a leaf to a creature asking for a gift leads to different responses, with each outcome generated live rather than pre-written.