实时世界模型进入“全科生”阶段，PixVerse R2先交卷！

TL;DR - AIsphere has launched PixVerse R2, a real-time world model designed to combine interactive, persistent scene generation with multimodal control. Its architecture separates general world-model capabilities from a distilled acceleration layer, aiming to preserve quality and consistency under real-time compute constraints.

• R2 unifies video, audio, actions, temporal context, and control signals within an Omni Causal AR framework.
• Dynamic chunks, multi-timescale memory, Hybrid Teacher/Diffusion Forcing, and an Error Bank target long-horizon drift, state discontinuities, and accumulated errors.
• A separate Real-Time Acceleration layer distills the pretrained model’s capabilities into a lower-latency system rather than constraining the base model from the outset.
• Demonstrated applications include prompt-controlled virtual worlds, interactive cinematic games, and real-time digital humans, though the article provides no standardized benchmark results.