Tsinghua team open-sources OpenMise, a two-stage Builder+Render pipeline fixing video models' spatial consistency failures

新智元 · wechat · 2026-11-09

SOTA video models like Seedance 2.5 still fail at detail consistency — chess pieces drift between frames or appear from nowhere even as characters, lighting and sets render correctly. Tsinghua and UPenn researchers open-sourced OpenMise, an agent-facing spatial scene editing tool that splits hard video generation into two easy stages.

Builder: a coding agent drives Blender (via MCP) to plan scenes, character paths, skeletal motion, camera keyframes and lights on one shared timeline, producing an inspectable graybox draft
Render: the draft, character references and prompts are fed to a video model like Seedance 2.5, which only needs to polish visuals instead of tracking every object's state

Key details: supports native Blender scenes, 3DGS scans and GLB/FBX assets with coordinate/scale calibration and spatial annotation; path checks account for character width so shoulders don't clip doorframes; cameras follow the character's traversed path to avoid wall clipping; errors can be fixed locally (adjust path, camera or lights) instead of re-rolling the whole clip.

Demo cases include walking a character from hallway to bedroom in a scanned 3DGS room with continuous over-the-shoulder tracking, delivering consistency that repeated sampling reportedly cannot match, and sharply cutting production costs for short-drama style content.

Original post →

Tsinghua team open-sources OpenMise, a two-stage Builder+Render pipeline fixing video models' spatial consistency failures

More from coding & agent