WorldDraft Benchmark Finds Top Video Models Leak Source Cues in Up to 64% of 4D Rendering Cases

WorldDraft, released Nov 13, puts 12 representative video models through 432 strictly paired 4D world-rendering tests spanning 48 executable world families and six dynamic challenge categories. Results show world preservation and visual realization in persistent tension: richer draft scaffolding anchors structure but hurts appearance fidelity, and conflicting source cues trigger source leakage in up to 64% of cases. The authors also ship WorldDraft-Primer, a plug-and-play visual prompting strategy that lifts both objectives on frozen video renderers.