WorldDraft

A benchmark for draft-to-video generative rendering

WorldDraftIf Code Builds the World,
Can Video Models Just Render It?

Preserve the world. Realize the appearance.

Paper Code Dataset Explore results
Paper Figure 1: performance of 12 video models across Rendering Performance, Granularity Gains, and Conflict Robustness.
12 models. Three complementary views of generative rendering.

The benchmark

One world, nine drafts.

We vary draft detail and source appearance while keeping the underlying world and target prompt fixed.

48 executable worlds6 challenge categories3 × 3 controlled conditions432 paired test cases
Six world challenges and the 3 by 3 factorial design crossing draft granularity with appearance alignment.
Abstract → Simulation → Previs   ·   Neutral / Aligned / Conflicting
From executable worlds to rendering evaluation
Benchmark workflow: executable worlds, controlled draft design, and rendering evaluation of world preservation, rendering fidelity, and appearance persistence.
Benchmark workflow · Figure 3, top panel

Exhibition board

Watch the world change.

Six cases from the paper. All 12 models, across every condition in the 3 × 3 grid.

Draft granularity
Appearance alignment
Target prompt

0.0 / 5.0 s

Shared playback · First 5 seconds · Full frames · Sound off Double-click a video to enlarge

Table 1

The leaderboard.

Rendering Performance, Granularity Gains, and Conflict Robustness contribute 50%, 25%, and 25% to Overall.

Download CSV ↓

WorldDraft results for twelve models. Activate a column heading to sort.

BestSecond best↑ Higher is better   ↓ Lower is better

Citation

@article{hou2026worlddraft,
      title={WorldDraft: If Code Builds the World, Can Video Models Just Render It?},
      author={Hou, Zhiyu and Shu, Wen-Jie and Chen, Harold Haodong and Ruan, Weilin and Liao, Chenfei and Zhang, Liuzhou and Lan, Disen and Xu, Xiaogang and Chen, Qi and Lim, Ser-Nam},
      journal={arXiv preprint arXiv:2611.16217},
      year={2026}
    }