Microsoft Open-Sources Code Video Model: Code the Scene First, Then Generate Video
xiaohu · x · 2026-10-13
Microsoft Research open-sourced Code Video Model, borrowing the Previs workflow from filmmaking:
- Code first, generate later: from a prompt, a coding agent (GPT-6 in the paper) writes an executable Three.js program that precisely places objects, trajectories, physics equations and camera moves, rendering a low-poly preview you can fix by editing one line of code.
- Four core capabilities: editable scene layouts and trajectories via coordinates; targeted local edits on the same scene (arch → doorway, humanoid → quadruped); independent camera vs. object motion planning; physics equations (spring, double pendulum) as motion reference.
- No retraining needed: works with off-the-shelf Minimax-H3 — the coarse animation only guides the first 25% of sampling steps, with reference images carrying realism afterward.
This adds an inspectable, editable intermediate step to video generation, mitigating the physics "black box" of pure text-to-video.
Related event: Microsoft Open-Sources Code Video Model for Controllable Video Generation(3 posts)→
More from Multimodal
- Magnific One Adds Brand Kit to Steer Image Generation With Your Brand Identity — LudovicCreator · 2026-10-14
- MIT CSAIL's InstructMesh makes genAI 3D models you can actually 3D print — MIT_CSAIL · 2026-10-14
- Black Forest Labs shares NVIDIA GTC talk links and Afterglow project — bfl_ai · 2026-10-14
- Black Forest Labs brings FLUX 3 demos and a keynote to GTC Berlin next week — bfl_ai · 2026-10-14
- Swing Choreo launches with a self-built motion benchmark: 61/100 vs 38 for the next best — andrew_n_carr · 2026-10-13
- No public benchmark for motion generation, so Swing built one: Choreo scores 61 vs 38 — andrew_n_carr · 2026-10-13