PixVerse R2: AIsphere’s Real-Time World Model, Explained

PixVerse R2 is AIsphere’s attempt to make AI video behave less like a finished clip and more like a world that keeps unfolding. Users can influence a running scene with text, images, audio and actions, while the model continues generating synchronized audiovisual output. AIsphere announced R2 on September 22, 2026; reports dated the formal launch in China to September 23. The company published a technical overview in August. 159

How users interact with an R2 world

Rather than issuing one prompt and receiving a fixed video, users can keep steering a session. PixVerse describes support for text, reference images, audio and action controls. Reports also describe keyboard-style WASD movement and prompts that can change a character, add an object or alter the environment while generation continues. The goal is for each new input to affect the current scene and what happens next—not simply start a disconnected clip. 16

That continuity is central to the “world model” idea: R2 is intended to use the scene’s history and current controls to generate the next synchronized audio-and-video segment. AIsphere says the model carries session events forward and aims to preserve relationships among characters, objects and spaces. 136

What the Zero Mark and Eve demos show

In Zero Mark, an interactive film-game created by Xiaolongbao, the same encounter can branch in response to a player’s choice. Offering a creature a dragon rather than a leaf produces a different live response, illustrating how an input can change the continuation of a scene. 12

A separate demo features Eve, a digital character. A report describes Eve responding to player questions and story developments using voice, persona, memory and relationship state, with the aim of maintaining a consistent identity over multiple exchanges. These are demonstrations of the intended interaction pattern, not independent evaluations of how reliably it works across longer sessions. 6

How the system aims to balance capability and speed

AIsphere describes R2 as having two main components. Omni Causal AR is the world-modeling component: it processes different input types and advances the scene over time, using the current state as context for what comes next. Real-Time Acceleration is intended to bring those capabilities into interactive operation. 1913

Other techniques address timing, continuity and computation:

• Dynamic chunking adjusts the length of generated segments to the input or event. Shorter chunks can support quick local responses; longer ones can represent more coherent motion or events. 3
• Multiple memory timescales are described as a way to keep both long-term anchors and recent changes in view. The technical overview identifies Sink Memory for persistent anchors, Rolling History for recent dynamics and Object KV Cache for reusing object-level representations. 10
• The Error Bank stores representative failure states and replays them during training, so the model can learn from generation errors rather than only from ordinary histories. 310
• Block-sparse attention is described as concentrating computation on selected, relevant relationships as the session grows. Secondary reports describe sparsity above 90% with little reported loss in output quality; that figure is a reported system result, not an independent benchmark. 716

What the performance claims establish—and what they don’t

AIsphere has reported a 35.8% reduction in visual drift across a long session in internal evaluations. That is a company-reported result, not an independently verified measure of performance across every use case. 2

The supplied sources do not provide a comparable independent latency, frame-rate or quality benchmark. So while PixVerse presents R2 as real-time and designed for longer-running consistency, the evidence here does not establish how it performs across different scenes, controls or extended sessions. The demos illustrate the concept; they do not settle how reliably it will work in broader use. 13

How R2 fits into PixVerse’s product direction

PixVerse introduced R1 in January 2026 as an earlier step toward continuously generated, interactive video; R2 is presented as an upgrade focused on broader inputs and longer-lasting scene state. The PixVerse Game Engine connects real-time generative video with AI agents and game mechanics, pointing toward experiences with roles, rules and changing world states. 1418

The available sources support describing that product direction, but do not provide a sufficiently grounded quotation to attribute a more specific vision for interactive worlds to CEO Wang Changhu.