Key Info

Qwen Developers open-sourced BlenderWorld, a benchmark of 223 tasks that tests whether visual agents can use Blender to turn text, image or video references into complete, editable 3D scenes. Evaluation is based on inspecting the actual project files and renders from multiple viewpoints rather than 2D appearance alone.

Highlights

  • 223 tasks covering text-, image- and video-conditioned 3D scene construction
  • Agents must produce editable Blender projects, not just images
  • Scoring uses scene inspection plus multi-view renders to check 3D consistency
  • Released on GitHub, ModelScope and Hugging Face, with showcases and support for plugging in custom 3D agents