Video models trade world preservation for target appearance
WorldDraft: If Code Builds the World, Can Video Models Just Render It? · arXiv:2611.16217
In plain words
The paper introduces WorldDraft, a set of tests for video models that receive a rough simulated scene and a text description. It checks two things separately: whether the video keeps the scene's motion and camera, and whether requested colors or materials appear on the right objects over time.
Models often do better at one than the other. The paper also introduces WorldDraft-Primer, which adds small visual cues to the first draft frame. On MiniMax-H3, it raises the measured target-appearance score, though this test does not show that every model benefits.
The research
A render that follows the prompt but changes the simulated action is not faithful, and preserving the action while ignoring the requested look is not enough either. These results make both checks relevant when judging whether a video model can render a supplied world.
WorldDraft tests video models as renderers, measuring whether they keep a simulated world's structure while applying and maintaining a prompt's visual attributes.
The benchmark pairs 48 executable worlds with three draft detail levels and three source-appearance conditions, then measures structure, target appearance, and whether that appearance persists. Across the full condition grid, evaluators found different strengths: SolarRenderer-33B scored higher than Wan 3.0 on world preservation, while Wan 3.0 scored higher on appearance fidelity. The split supports a trade-off, not a single overall winner.
The scores average results across simulated worlds and input conditions; they do not describe every case. The reported WorldDraft-Primer comparison is for MiniMax-H3 and does not show the same gain for every renderer.
What the evidence shows
- Two jobs in one render
WorldDraft is a benchmark that tests video models as renderers: given a simulated scene draft and a text prompt, can they keep the world's structure while producing the requested look? The test treats these as separate requirements.
A model may preserve motion and camera while missing target colors, or match the requested appearance while changing the simulated world.
- Different models, different strengths
The comparison separates two capabilities rather than ranking an overall winner: SolarRenderer-33B does better at keeping the simulated world intact, while Wan 3.0 better matches requested visual attributes. Their averaged scores show a trade-off on WorldDraft, not that either model leads on every case.
- More detail can shift the balance
Paired tests show how the inputs relate to that tension. With neutral source appearance, moving from an abstract draft to a detailed preview generally improves structural preservation but often reduces appearance fidelity and binding; Seedance 2.5 is an exception, improving on both appearance measures.
In a separate test with conflicting source and target colors, contradictory source colors often remain on target objects. The findings point to difficulty both replacing appearance cues and keeping attributes attached to the right objects.
- First-frame cues help one model
Compared with unmodified MiniMax-H3, the added first-frame cues improve both target-appearance fidelity and the composite score on WorldDraft. This supports a gain for this model and test, while the paper's noted failures on weaker models and dense previews limit a broader conclusion.
Deep
- Original paper authors
- Zhiyu Hou, Wen-Jie Shu, Harold Haodong Chen, Weilin Ruan, Chenfei Liao
- Submitted to arXiv
- License
- CC BY 4.0