EVENT DOSSIER
On September 7, 2026, Hugging Face launched the unified training-free video editing framework called EditVid. This framework supports various editing paradigms, including instruction-guided and reference-guided approaches, and employs technologies such as sparse causal memory, relational-based post-attention token injection, and soft latent mixing. In tests on the FiVE dataset, its accuracy reached 78.16%, significantly outperforming the strongest training-free baseline (58.95%). User surveys showed that EditVid led seven competing methods with an overall preference rate of 51.8%.
Hugging Face 发布名为 EditVid 的统一训练免费视频编辑框架，支持指令引导与参考引导等多种编辑范式。该框架结合稀疏因果记忆、基于对应关系的后注意力令牌注入及软潜在混合技术，在 FiVE 数据集上取得 78.16 的准确率，显著优于最强训练免费基线（58.95），并在用户研究中以 51.8% 的整体偏好率领先七种竞争方法。