TL;DR of the paper, generated automatically.

A benchmark asks if video models can render code-built 4D worlds.

• Tradeoff: keeping motion and camera paths clashes with matching target looks.
• Drafts: richer scaffolding anchors structure but impairs appearance fidelity.
• Leakage: conflicting cues cause source leakage, up to 64%.
• Fix: WorldDraft-Primer, a visual prompt, consistently improves both goals.