AIsphere, the developer of the PixVerse video generation platform, has released PixVerse R2. This model is a universal real-time world system that creates continuous interactive audiovisual worlds instead of generating fixed clips.
R2 was unveiled on September 23rd and follows version R1, which the company announced in January. Users can now open demonstration spaces for games, interactive films, and digital humans on the website world.pixverse.video.
With R2, users can create worlds based on text descriptions or images. They can then control a character using WASD keys, change the camera angle, and input prompts to replace characters, add objects, or alter the environment while the scene continues to run. The company emphasizes that previously entered data is preserved: any action or prompt updates the world's state and determines subsequent developments, rather than only affecting the current moment.
In the interactive game film Zero Mark, created by author Xiaolongbao, an entity blocks the player's path and demands a gift; choosing a dragon or a leaf leads to different live-generated outcomes. Another author, Jade Wu, developed a story around a digital character named Eve, who maintains a constant identity and memory throughout numerous interactions.
AIsphere positions R2 as a solution to a long-standing compromise: typically, real-time generation requires a small and fast model, whereas broad capabilities demand a larger one. R2 divides this task into two levels. The base Omni Causal AR model, which is a causal autoregressive core, trains continuously and assimilates short and long videos, multimodal links, audio, and action controls, scaling across five dimensions: model, data, tasks, control signals, and time horizons.
Furthermore, the real-time acceleration layer takes this same base model and transforms it into an ultra-fast version that operates in real-time, instead of training a separate fast model. Auxiliary technologies include dynamic chunking for processing signals operating at different time scales, three memory channels for preserving world rules, recent movement and object states, and an Error Bank that reproduces the model's own failures during training.
AIsphere reports that the Error Bank reduced the long-horizon brightness drift metric by 35.8% in internal tests, while sparse attention maintained quality almost unchanged at sparsity levels above 90%.
The company is candid about its limitations: under strict real-time and latency constraints, the quality of R2's single-pass output still lags behind leading offline video models. The PixVerse game engine, first launched in July, now runs on R2. Founder and CEO Wang Changhu believes that real-time interactive video is not the entire scope of world modeling, but it is the earliest path through which people can gain experience. He expects R2 to transform from a content creation tool into an environment users can enter at any time.