qwenlm/blenderworld is a benchmark that evaluates visual agents building editable 3D scenes in Blender from real-world references. The Qwen team released it as a GitHub repository, and no separate latest release version was given.
The benchmark consists of 223 tasks and 2,651 binary grading criteria. The inputs comprise 78 texts, 130 images, and 15 videos, and 59 tasks come with specification files such as dimensions or materials. The agent receives a task description and references, then submits final.blend along with the requested deliverables. The grader renders the scene with the Cycles renderer at 1024x1024 resolution and 64 samples from eight viewpoints chosen to match the situation, and allows up to two corrective renders per viewpoint. Structural evaluation is done by agent-based review, while the visual criteria and the eight quality ratings are evaluated by a vision-language model. The final score gives half weight to the weighted structural score and half to the average of the eight quality ratings.
The published grader imposes no score cap. It is also hard to rule out evaluator bias, because GPT-5.6-terra is used in all three roles of structural judging and visual evaluation, and the repository has only one test path, harbor_assets/test.sh, which limits automated verification of the grading logic itself. Score comparisons therefore carry the limitations of the evaluation method and the scope of verification.