The authors built BlenderWorld, the first comprehensive benchmark for evaluating visual agents on practical 3D creation workflows in Blender.
Benchmark shows AI agents struggle with fine-grained 3D spatial coherence in Blender
BlenderWorld evaluates 13 frontier multimodal models on 223 realistic 3D creation tasks; the top model scores 79.9 but fully satisfies only 42.2% of tasks.
Chinese Tech
Xinyu Fang · Yuchi Wang · Weihao Yu · Rongyao Fang · Zhaohai Li · Zhixun Li · +9 more
Zhejiang University · Alibaba Group · The Chinese University of Hong Kong · Fudan University · National University of Singapore
Research Digest··3 min read
The authors introduce BlenderWorld, a benchmark of 223 realistic 3D creation tasks in Blender, spanning reconstruction, modification, and design from text, image, video, and structured specification inputs.
§