EditWorld adds precise editing and flexible referencing to video world models, scoring 80.0 on editing

NanyangTechnologicalUniversity · hf · 2026-09-29

NTU's EditWorld extends video world models from exploration to precise modification, streaming editing instructions and reference images during autoregressive generation. Key pieces: Gated Causal Attention for temporally varying editing conditions, a Sparse Context mechanism for bounded long-horizon context, joint autoregressive + bidirectional training with annealed self-resampling, and a data synthesis pipeline for editing supervision. It also introduces WBench-Editing, where EditWorld achieves 73.8 overall and 80.0 editing scores, substantially outperforming existing methods. Code is open-sourced.

Original post →

More from Multimodal

Multimodal channel →