Precise Editing and Flexible Referencing for Interactable Worlds
EditWorld is built to change existing generated worlds, not just let users move through them.
The paper says the model streams editing instructions and reference images during autoregressive video generation. It adds Gated Causal Attention for changing edit conditions over time and Sparse Context for longer inference with bounded history. The authors also introduce WBench-Editing, where EditWorld reports a 73.8 overall score and 80.0 editing score, ahead of existing methods on editing metrics. Source: HF Daily Papers' note
The paper says the model streams editing instructions and reference images during autoregressive video generation. It adds Gated Causal Attention for changing edit conditions over time and Sparse Context for longer inference with bounded history. The authors also introduce WBench-Editing, where EditWorld reports a 73.8 overall score and 80.0 editing score, ahead of existing methods on editing metrics. Source: HF Daily Papers' note
score 5