Reproduce “BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinyu Fang, Yuchi Wang, Weihao Yu, Rongyao Fang, Zhaohai Li, Zhixun Li, Tiancheng Gu, Xilin Wei and 7 more
BlenderWorld evaluates whether AI agents can produce complete, editable Blender projects for varied 3D briefs, not only convincing images. It combines task-specific views and visual ratings with checks of project structure and requirements. Across 223 tasks, GPT-6-Astra has the highest mean fused score, 79.9/100, but passes every rubric criterion on only 42.2% of tasks, under the benchmark’s one-run protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.