WorldDraft, a preprint submitted to arXiv on November 12, 2026, tests video models as renderers of programmed scenes. The authors report trade-offs between preserving motion and applying requested appearances. arXiv records the submission date.
The authors’ repository separately dates the benchmark data and evaluation-code release to November 11 and the paper release to November 13. These describe different steps in the release, rather than a single launch date.
What WorldDraft measures
The benchmark contains 432 paired cases across 48 executable world families. Each family has nine conditions: three levels of draft detail crossed with neutral, matching or conflicting source appearances. Dynamics and the target prompt stay fixed. The six challenges include multiple interacting entities, staged dynamics, occlusion and reappearance, camera movement, deformation and hybrid interactions, according to the repository.
This is a synthetic benchmark: the test structure is deliberately controlled. The following distinctions in the authors’ evaluation separate keeping the scene intact from making it look right.
| Evaluation dimension | What it checks |
|---|---|
| World preservation | Whether entities, spatial arrangements, motion, events and camera movement follow the draft. |
| Rendering fidelity | Whether the generated video shows the target attributes specified in text. |
| Appearance persistence | Whether correct target attributes remain attached to the same entities over time. |
The repository lists the evaluation metrics. The project site presents synchronized video comparisons and a downloadable results table. Its overall ranking combines rendering performance, granularity gains and conflict robustness with weights of 50%, 25% and 25%, respectively; it is a composite of different questions.
What the authors found across 12 models
The authors report that richer drafts often helped structure while hurting appearance. Seedance 2.5 was an exception, with positive appearance and binding changes.
Wan2.1-VACE recorded 64.17% source-color leakage under conflicting appearance at simulation-level detail. That figure applies to this model and condition. These are findings from the preprint’s experiments.
Primer changes the first frame
WorldDraft-Primer adds object and appearance cues to the first frame without renderer retraining. The authors report gains on OmniWeaving and MiniMax-H3. This two-model experiment has a narrower scope than the benchmark.
Why the release matters for university research
For university groups studying simulation-to-video workflows, the release offers material for testing the rendering stage. The dataset card describes draft videos, target prompts, exported ground truth and Blender scenes; the repository provides download and evaluation instructions. The 432 figure describes benchmark cases, rather than repository files or dataset-viewer rows.
The research-office implication is about the evidence behind a demonstration: visual appeal, adherence to a programmed world and persistence of requested attributes answer different questions. The distinction connects with CASRAI’s coverage of programmatic video evaluation and its definition of a reproducible AI experiment. WorldDraft supplies a framework for investigating these questions; its preprint results alone do not establish suitability for scientific simulation or institutional deployment.
Sources
- Hou and colleagues, WorldDraft arXiv submission record and preprint, submitted 2026-11-12. Checked 2026-11-15.
- WorldDraft authors, official repository and release notes. Checked 2026-11-15.
- WorldDraft authors, project comparisons and leaderboard. Checked 2026-11-15.
- KevinHuge, WorldDraft dataset card. Checked 2026-11-15.







