Microsoft Research's Code Video Model generates video from editable code, scoring 85.82 vs VACE's 80.59
jiqizhixin · x · 2026-11-07
Microsoft Research's Code Video Model technical report treats a video shot as an editable program: a coding agent builds a Three.js scene specifying objects, trajectories, camera motion and timing, which renders a proxy video for review before final generation, with a reference image controlling appearance.
- Proxy conditioning runs only for the first 5 of 20 sampling steps in the MiniMax-H3 setup, while text/image guidance persists throughout—avoiding the preview's flat rendering style without extra video-model training
- On CodeVideoBench, the method scores 85.82 under a GPT-6 evaluator versus VACE's 80.59
The result: creators can edit layout and motion like code before committing to final video generation.
More from Multimodal
- Grok Bot + Pokee Isaac: Four Lyric Videos in Under an Hour — Kyrannio · 2026-11-07
- Viral "Time travel" AI video shows a man standing before the Taj Mahal at sunset — Sam77yusuf · 2026-11-07
- Third-party Animator Grok Bot turns one sentence into a multi-scene animated video — nima_owji · 2026-11-07
- Gemini + MCP driving invideo editor turns raw drift footage into a finished edit — Neat_Bit1936 · 2026-11-07
- Claude's multi-agent pipeline built an RTS game intro video, rendered in ~3 hours on rented 8xA40 — Liqhthouse · 2026-11-07
- Recraft Studio October update: Logo Maker, Vary, Visual Annotations and six new models — aziz4ai · 2026-11-07