PixVerse launches R2 real-time world model
PixVerse has released R2, an upgrade to its real-time world model that keeps generated worlds coherent for longer and carries user input forward within a session, so the worlds behave more like playable games than one-off clips.
PixVerse announced R2, the successor to R1. The company launched R1 in January 2026 and describes it as the first real-time world model. A real-time world model doesn't return a fixed video clip. It generates a navigable world that keeps running while new input arrives.
According to PixVerse, R2 stays coherent over "far longer sessions," keeps track of what happens during a session, and accepts text, reference, audio and action-control inputs in the same running world. The main change is that input persists. A prompt or action updates the world's state and affects what happens later, rather than changing only the current frame.
The company pointed to two demos built by creators. In Zero Mark, made by creator Xiaolongbao, offering a creature a dragon or a leaf leads to different story branches, each generated live. In a story built by Jade Wu, a character named Eve keeps the same identity and memory across a conversation. In general play, users move a character with WASD keys and change the world with prompts.
On architecture, PixVerse says R2 replaces a multi-stage training pipeline with two stages built on one foundation. "Omni Causal AR" is a single causal model that keeps training on more data, input types and tasks. An acceleration layer then compresses that same model to run in real time, so it doesn't need a separately trained fast model.
The announcement gives no latency figures, session lengths or benchmark results. PixVerse says those details are in a separate R2 technical report.