R2 wants generated worlds to last beyond a few seconds

PixVerse is extending its real-time World Model with R2, which can maintain session state, respond to player actions, and incorporate text, audio, references, and controls while the world continues to be generated.

A world that no longer starts over

A generated video ends with its final frame. A real-time World Model faces a different problem: it has to keep producing the environment while the user moves through it, takes actions, or provides new instructions.

With R2, PixVerse is primarily extending that continuity. An action taken earlier in a session can continue to influence characters, environments, or the narrative later on. The system maintains an active world state rather than treating each new interaction independently.

Inputs can combine text, references, audio, and action commands. In one demonstration, the player moves with WASD controls, operates the camera, and uses prompts to alter the environment without interrupting generation. Subsequent interactions build on what has already happened. The story develops while you play

That memory also affects character behavior. PixVerse demonstrates Zero Mark, an experience built with R2 where a creature blocks the player's path and asks for a gift. Giving it a dragon or a leaf produces different reactions generated during the session rather than two entirely pre-scripted branches.

Another demonstration features Eve, a character that maintains her identity and conversational context across multiple exchanges. Her responses, expressions, and movements evolve according to both the conversation and the player's progress through the story.

The distinction from a conventional game lies partly in what is computed beforehand. Characters, events, and consequences can be produced during the experience instead of being completely defined before runtime. Two stages to maintain real-time generation

Keeping that process running creates additional constraints. The model has to preserve characters and environments, process new commands while running, remember previous events, and correct small deviations before they accumulate, all while keeping latency low enough for interaction.

R2 splits that work into two components. Omni Causal AR provides the capability foundation: a causal model continuously trained across more data, tasks, modalities, and longer time horizons. A second layer then compresses that same model for real-time execution.

PixVerse contrasts this architecture with pipelines where several models or training stages hand work off to one another. Here, the faster runtime version derives from the same underlying foundation rather than being trained separately from scratch. From World Model to game engine

R2 now powers the PixVerse Game Engine, with early experiences available to the public. The engine is designed for environments where creation and interaction happen at the same time.

That shifts part of the creation process into runtime. Instead of preparing every character, event, and branching path before the player enters the experience, the system can generate some of them in response to incoming actions.

A selection of R2 worlds can currently be explored through PixVerse World.