Qwen open-sources BlenderWorld, a 223-task benchmark testing if agents can build editable 3D scenes
petewoodbridge · x · 2026-10-23
Qwen's team open-sourced BlenderWorld, a benchmark built on the observation that looking right in 2D doesn't mean a scene holds together in 3D.
- 223 tasks where agents use Blender to build editable 3D scenes from text, image, or video references
- Evaluation goes beyond renders: actual projects are inspected via scene inspection and multi-viewpoint renders
- Core question: can a visual agent turn a real-world brief into a complete, editable 3D scene?
The benchmark is available for download, and developers can plug in their own 3D agents to compare results.
More from Research
- GenReasoning team reflects on six-month grind: lean startup, zero paid promotion — rosstaylor90 · 2026-10-23
- 33K-Star GitHub Repo ML-From-Scratch Rebuilds ML Algorithms in Pure Python — techNmak · 2026-10-23
- THU-KEG Releases RIDAR to Detect and Intervene in Research Agent Integrity Violations — THU-KEG · 2026-10-23
- Economist proposes AI-first journals: automated review, one-click replication, formal proofs — Afinetheorem · 2026-10-23
- Economist proposes AI-first journals: AI first-pass review, 1-click replication, formal proofs — Afinetheorem · 2026-10-23
- Same LoRA on every Qwen3.5 size: 9B beats 27B, and nearly all the gain is learning to say no — spilldahill · 2026-10-23