WorldDraft Evaluates Whether Video Models Can Generatively Render Code-Built 4D Worlds

Written by

in

WorldDraft introduces a benchmark for evaluating the generative rendering capabilities of video models, while WorldDraft-Primer improves the world consistency and visual fidelity of frozen models. The benchmark includes 432 strictly paired test cases from 48 families of executable 4D worlds, covering six types of dynamic challenges and using a 3×3 factorial design to combine draft granularity with appearance-alignment conditions.

Source:Hugging Face 社区论文精选 · Read Original

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *