WorldPlay: Real‑Time Interactive World Modeling with Long‑Term Geometry Consistency

World models aim to enable agents to perceive and interact with dynamic 3‑D environments, but achieving low‑latency real‑time generation while preserving long‑term geometric consistency remains an open challenge.

WorldPlay Overview

The Tencent HyperAI (混元) team released WorldPlay, an open‑source system that delivers real‑time interactive world modeling with sustained geometric consistency, addressing the speed‑vs‑memory trade‑off of prior methods.

Key Technical Innovations

Dual‑Action Representation : Captures both keyboard and mouse inputs through a dual‑action scheme, ensuring accurate and stable interactive control.

Reconstructed Context Memory Mechanism : Introduces a dynamic memory module that rebuilds contextual information from historical frames and applies a temporal reconstruction strategy, allowing access to geometrically important but temporally distant frames and mitigating memory decay.

Context‑Forcing Distillation : Aligns the memory context of teacher and student models so the student retains long‑range information while maintaining real‑time inference speed, thereby suppressing error drift.

WorldPlay can stably generate long video sequences at 24 FPS and stream 720p high‑definition video, outperforming existing techniques on multiple metrics and demonstrating strong generalization across diverse scenarios.

Online demo: https://go.hyper.ai/Dgd3Z