Precise Editing and Flexible Referencing for Interactable Worlds

2026-09-28

AI Summary

This study addresses the limitation of existing video world models, which support only exploratory navigation and lack precise editing capabilities over generated content. To this end, this work proposes EditWorld, a framework that enables precise editing of interactive worlds through streaming editing instructions and reference images. Methodologically, it introduces gated causal attention to handle time-varying conditions and designs a sparse context mechanism to maintain long-horizon reasoning. The approach further integrates autoregressive and bidirectional training, an annealed self-resampling technique, and a dedicated data synthesis and annotation pipeline. Experimental results demonstrate that EditWorld achieves an overall score of 73.8 and an editing score of 80.0 on the WBench-Editing benchmark, significantly outperforming existing methods.

Abstract

We present EditWorld, a video world model for precise editing and flexible referencing in interactable worlds. Existing video world models primarily focus on navigation, letting users explore generated worlds but offering limited control over how existing world content is modified. EditWorld extends world modeling from exploration to precise modification by streaming editing instructions and reference images during autoregressive generation. To support these capabilities, EditWorld introduces Gated Causal Attention for temporally varying editing conditions and reference images, together with a Sparse Context mechanism that maintains a bounded historical context for long-horizon inference. We further adopt joint autoregressive and bidirectional training with annealed self-resampling, and construct a dedicated data synthesis and annotation pipeline that provides supervision for world editing. We also present WBench-Editing to systematically evaluate streaming world editing capabilities. EditWorld achieves the best overall performance on WBench-Editing with an overall score of 73.8 and an editing score of 80.0, substantially outperforming existing methods on editing-related metrics. https://github.com/leoisufa/EditWorld

Problem

• video world model
• precise editing
• flexible referencing
• interactable worlds
• world modification

Innovation

• Video World Model
• Gated Causal Attention
• Sparse Context Mechanism
• Streaming Editing
• Autoregressive Generation