BlenderWorld: Benchmarking Visual Agents for 3D Creation from Real-World References

BlenderWorld. Benchmarking Visual Agents for 3D Creation from Real-World References. By Xinyu Fang and fourteen others, from Zhejiang University, Alibaba Group, and three other institutions. Published on the fourteenth of October. Its code repository is public and real: about twenty source files.

Multimodal large language models can now act as autonomous agents that build and edit 3D models, either by writing code or by manipulating the software directly. In real workflows that is a demanding job. The model has to infer the shapes of objects, their materials, how things are laid out in space, and how the scene is lit, all from references that come in many forms. And it has to follow instructions strictly, whether the request is to reconstruct something, modify it, or design it from scratch. The authors argue that existing benchmarks cover only a narrow slice of these open-ended requests across real-world scenarios.

BlenderWorld is presented as the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation. It asks a plain question: handed a real-world reference and an instruction, can a model produce a Blender project that matches what was asked for?

The benchmark has two hundred and twenty-three tasks. They span architecture, products, natural environments, and scientific visualization, and they are driven by four kinds of input: text, images, video, and structured specifications.

Scoring a deliverable like that is not straightforward, so the authors set up a unified evaluation protocol. Each submitted Blender project is rendered from multiple viewpoints. Those renders are checked against two thousand six hundred and fifty-one task-specific binary rubrics, and they also receive holistic scores for visual quality. The rubrics and the holistic scores are combined into a fused score.

Thirteen frontier models were benchmarked. GPT-6-Astra reached the highest overall fused score, at seventy-nine point nine. Opus-5 came next with seventy-three point zero, and GPT-5.6-Sol followed at sixty-seven point seven.

Behind the ranking sits a common pattern. Current models can parse broad instructions, but they still struggle badly with fine-grained 3D spatial coherence, and with reproducing structure accurately inside Blender. The authors read this as a critical bottleneck.

A top score near eighty could be heard as a task close to finished. The findings say otherwise: models still struggle with where things sit relative to one another, and whether the built structure matches what was asked for. That leaves substantial headroom, in the authors' words, for future 3D-native agentics.