EditWorld Moves Video World Models from Exploration to Precise Editing

Introduction

Video world models are often described as systems that generate environments users can explore and interact with. Yet exploration is only one side of interaction. In many practical settings, users also need to change what is already present: they may want to alter an object, introduce a visual reference, or revise a scene after generation has begun. Existing models have generally provided less precise control for this kind of modification.

EditWorld, presented by researchers from Nanyang Technological University and other institutions, targets this gap. It frames world modeling not only as the prediction of future observations, but also as the controlled modification of an evolving interactive world.

Key ideas

• Editing is streamed into generation. EditWorld allows editing instructions and reference images to arrive during autoregressive generation. This makes the model suitable for an iterative workflow in which a user observes the current scene and issues new instructions without restarting the entire process.
• Gated causal attention handles changing conditions. Editing signals are not necessarily relevant for every moment of a video. The proposed Gated Causal Attention is designed for temporally varying instructions and reference images, allowing the model to regulate how these conditions influence generation while preserving causal temporal modeling.
• Sparse context supports longer rollouts. As an interactive world is generated over a longer horizon, retaining every past state can make inference increasingly expensive. EditWorld’s Sparse Context mechanism keeps a bounded historical context, aiming to preserve useful world information without allowing the context to grow without limit.
• Training is adapted to editing behavior. The method combines autoregressive and bidirectional training with annealed self-resampling. The paper also describes a dedicated synthesis and annotation pipeline that supplies supervision specifically for world-editing tasks.
• Evaluation emphasizes controllability. WBench-Editing is introduced as a benchmark for streaming world editing. According to the paper’s abstract, EditWorld obtains an overall score of 73.8 and an editing score of 80.0, with particularly strong results on editing-related measures.

Why it matters

The central contribution of EditWorld is a shift in the role of a world model. Instead of treating generation as a one-time process followed by passive exploration, it treats the world as an evolving object that can be revised while it is being generated. This is closer to how people work with creative tools: inspect an output, provide a correction, add a visual reference, and continue from the updated state.

That capability could matter for interactive storytelling, game prototyping, virtual environments, and visual production. Still, the available material does not establish how EditWorld behaves in every challenging setting. Long-term consistency, conflicting instructions, reference-image diversity, inference cost, and performance outside the benchmark remain important questions. The paper’s results nevertheless point to a broader direction for world models: being explorable is useful, but being precisely editable may be what makes them practical.

Hugging Face Daily Papers