Tsinghua and Penn open-source OpenMise for consistent AI video
Teams from Tsinghua University and the University of Pennsylvania have open-sourced OpenMise, a spatial-scene editing tool that uses coding agents and Blender to build editable 3D plans before an existing video model renders the final footage. The two-stage Builder-and-Render workflow separates motion, camera, lighting, and object-state planning from visual rendering, with chess and indoor 3DGS demonstrations reported by New Zhiyuan. The examples showed better consistency than direct Seedance 2.5 generation, but these comparisons are team demonstrations rather than independent measurements.
OpenMise offers a practical way to improve spatial, motion, and camera continuity in AI video without requiring a new video model, while preserving editable Blender projects and structured plans for targeted revisions.
新智元 2026-11-09 12:56 北京
新智元报道
目前最强的视频生成大模型,例如seedance 2.5在根据用户意图生成视频的时候有一些明显的缺点。
以seedance 2.5创作两个指定角色对弈国际象棋作为例子,可以发现sota视频生成模型在细节的一致性上仍然有很多不足。
提示词:生成两个人面对面下国际象棋的写实视频,完整展示西班牙开局,并严格依次执行 e4 e5、Nf3 Nc6、Bb5 a6,直到后续的王车易位、棋子移动与对手回应。要求棋步合法,棋盘状态连续,不得出现多余棋子、错误落点或棋子凭空消失。
#OpenMise#Tsinghua University#University of Pennsylvania#Blender#Seedance 2.5#DeepSeek V4.1 Flash#KIRI