What it is
Benchmark for generating executable 3D worlds from text and images, scored on whether the constructed world actually works rather than just looks right. 1K-10K benchmark instances organized into three task sets (T1/T2/T3) with task files, assets, reference images/frames, API prompts, and judge prompts for code-based 3D world construction in Blender.
Key results
- 1K-10K task instances
- 3 task groups: T1, T2, T3
- Text, images, frames, assets, and eval prompts
Useful for benchmarking models that generate executable 3D world-building code from visual and text inputs.