Skip to content

Source paper

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling

Wenqiang Sun, Haiyu Zhang, Haoyuan Wang and 7 more

Wenqiang Sun, Haiyu Zhang, Haoyuan Wang, Junta Wu, Zehan Wang, Zhenwei Wang, Yunhong Wang, Jun Zhang, Tengfei Wang, Chunchao Guo

arXiv 2512.14614 · 202561 upvotes on Hugging Face

Research ideas from this paper (4)

Generated by AI. The paper’s authors did not write these.

Multi-Entity Action-State Modeling for Interactive Environments

Extend the action representation into a multi-token descriptor with persistent identity latents and a cross-attention interaction layer to model multi-agent collisions and collaborative behaviors.

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling
🔴Hard🔬Significant ResearchMultiple GPUsNo close match found#Robotics#Generative media

Physically-Consistent Latent Dynamics via Differentiable Constraints

Integrate a differentiable physics engine module into video latent spaces and apply a physics-consistency loss during memory training to enforce Newtonian motion priors.

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling
🔴Hard🔬Significant ResearchLarge clusterRelated work found#Generative media#Robotics

Hierarchical Sparse Memory for Ultra-Long Horizon Consistency

Develop a hierarchical memory retrieval system using a global-local cache and KV-cache pruning to maintain constant attention overhead during ultra-long horizon video generation.

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling
🟠Medium💡Moderate InnovationMultiple GPUsRelated work found#Retrieval & memory#Efficiency

High-Dimensional Articulated Control in Latent World Models

Implement a skeletal-aware action encoder and pose-conditioned VAE to map high-dimensional joint angles and velocities to visual environment latents for articulated control.

WorldPlay: Towards Long-Term Geometric Consistency for Real-Time Interactive World Modeling
🔴Hard💡Moderate InnovationMultiple GPUsNo close match found+1#Robotics#Generative media

AI summary

WorldPlay is a streaming video diffusion model designed for real-time, interactive world modeling with long-term geometric consistency. It resolves the trade-off between generation speed and memory preservation by introducing a dual-action control representation, a reconstituted context memory that reframes past frames, and a novel distillation method called Context Forcing.

How it works (AI summary)

The architecture employs a Diffusion Transformer (DiT) finetuned for chunk-wise autoregressive generation. It utilizes 'Temporal Reframing' to rewrite positional embeddings of retrieved past frames to maintain influence over long distances, and 'Context Forcing' to align the memory distributions of a bidirectional teacher and an autoregressive student during real-time distillation.