WorldDraft benchmark: can video models just render worlds built by code?
Zhiyu Hou · hf · 2026-11-13
Researchers propose WorldDraft, a benchmark probing a new division of labor where coding agents build 4D worlds and video models act as generative renderers. A reliable renderer must satisfy world preservation (keeping prescribed 4D kinematics and camera paths) and visual realization (persistently binding text-specified attributes).
- 432 strictly paired test cases across 48 executable 4D world families and 6 dynamic challenge categories, with a 3×3 factorial design crossing draft granularity and appearance alignment.
- Evaluations of 12 models show world preservation and visual realization are in stark tension; richer draft scaffolding anchors structure but impairs appearance fidelity; conflicting source cues trigger source leakage up to 64%.
- WorldDraft-Primer, a plug-and-play visual prompting strategy, consistently improves both objectives on frozen video renderers.
More from Multimodal
- Minimax Image impresses: the team still ships open models — Druck_Triver · 2026-11-13
- KREA2 vs Z-Image vs Moss vs Qwen 2.1: which local image model to pick — tj-tj-tj-tj · 2026-11-13
- ShowTake launches: coding agents that also film your app demo videos — Scobleizer · 2026-11-13
- LoRA training help: white-background datasets leak shading into scenes — sonsuka · 2026-11-13
- DAWN harmonisation model now works with Gaussian splats for consistent relit footage — Scobleizer · 2026-11-13
- BirdDreams driving world model generates perspective + BEV video from a single image — lealtaixe · 2026-11-13