WorldDraft: If Code Builds the World, Can Video Models Just Render It?
More detailed world drafts helped video models preserve motion and structure but made the requested appearance harder to achieve.
WorldDraft tests whether video models can render code-built worlds while preserving prescribed motion and camera paths and applying text-specified visual attributes. Its benchmark covers 432 paired cases across 48 executable world families, with evaluations of 12 models. Conflicting visual cues in the drafts caused source appearance to leak into outputs at rates up to 64%. The authors report that WorldDraft-Primer, a visual prompting strategy requiring no model updates, consistently improved both world preservation and visual realization.
HF Daily Papers' note
WorldDraft tests whether video models can render code-built worlds while preserving prescribed motion and camera paths and applying text-specified visual attributes. Its benchmark covers 432 paired cases across 48 executable world families, with evaluations of 12 models. Conflicting visual cues in the drafts caused source appearance to leak into outputs at rates up to 64%. The authors report that WorldDraft-Primer, a visual prompting strategy requiring no model updates, consistently improved both world preservation and visual realization.
HF Daily Papers' note
score 5