Code Video Model uses code as the engine for controllable, physics-grounded video generation
angrypenguinPNG · x · 2026-10-12
Researchers introduced Code Video Model, bringing the film-production previs concept to video generation: agentic coding as the engine for controllable video, with paper, code, and data released.
- Pipeline: a coding agent turns a prompt into an executable program that lays out the scene, animates objects, places cameras, and computes dynamics from equations; rendering yields a sub-minute proxy video plus a reference image — grounded and precisely editable via code.
- Key trick: training-free early-step conditioning — a pretrained video model attends to the proxy only during early sampling steps (where structure forms), then drops it, avoiding the flat-appearance copying problem while adopting the reference image's look.
- Applications: physics-grounded motion, bullet time, gaming, scene editing, world-model data, 3D/4D reconstruction, robotics simulation.
- They also release CodeVideoBench for evaluating controllable video generation.
More from Multimodal
- Minimax H3 faces go blurry in wide shots: ComfyUI users seek fixes — Unique_Bag1444 · 2026-10-12
- Fixing silhouette and palette, exploring only costume: scoping a local SDXL character workflow — Alchemistwiza · 2026-10-12
- LoRA Pilot ep.3: character dataset prep with browser upload, image review and captions — no3us · 2026-10-12
- vivo and Xidian Propose StereoBind for Spatially Consistent Stereo Audio in Video Generation — Hanmo Chen · 2026-10-12
- MiniMax voice cloning trouble: cloned character babbles and speaks too early — Nokersoek · 2026-10-12
- Dev building an AI DJ desktop app that matches BPM/key and fixes AI music drift — AIandDesign · 2026-10-12