H3-World’s Action Control: Why One Prompt Is Not Enough
An analysis in 掘金 · 人工智能 examines how H3-World binds actions to specific video intervals through packed sequences, temporal coordinates, and attention masks, based on the paper’s version 1 and pinned official code. The article includes a Python scalar checker that validates a simplified visibility matrix, span constraints, padding behavior, and a deliberately swapped action mapping. It does not run the model, download weights, measure device latency, or independently verify successful video generation.
The article identifies how action timing is represented and routed in H3-World, while making clear that its static checks and paper examples do not establish successful generation or real-time performance.
H3-World 的动作控制,为什么不能只靠一句 提示词 ?
作者静态机制/未执行记录;不代表模型生成成功或实时测量。
给视频模型写“镜头先向左转,再向右转”,听起来已经把动作和顺序交代清楚了。但模型读懂左右,不等于这两个方向分别进入了指定的视频时间段。H3-World 要解决的就是这个差别:动作仍用语言表达,时间归属却不能只留在一句全局描述里。
本文以 2026 年 10 月 12 日核对的论文 v1 与官方固定代码为准。讨论对象是按动作生成固定长度视频片段的接口,不是实时游戏控制器。没有运行模型、下载权重或测量设备延迟。逐帧动作如何压入时间槽,是进入模型前需要核对的另一处边界。这里假定槽级动作已正确准备,继续追问它们如何进入主干,以及读代码时怎样证明“左句属于左段、右句属于右段”。