A native project
The agent receives the brief, references and attachments, then creates an editable Blender project. Evaluation rubrics stay hidden during generation.
A benchmark for visual creation agents
BlenderWorld.
Benchmarking Visual Agents for 3D Creation
from Real-World References
1Zhejiang University2Alibaba Token Hub, Alibaba Group3The Chinese University of Hong Kong4Fudan University5National University of Singapore
Real tasks. Editable scenes. Observable results.Step inside the artifacts
223real-world tasks
13frontier models
9creative domains
2,651binary rubric criteria
Meet BlenderWorld.
English · 1 min 11 sec
01 The benchmark
Seeing is only
the beginning.
Can a visual agent turn a real-world brief into a complete, editable 3D scene?
BlenderWorld asks agents to interpret text, images and videos, work inside Blender, and deliver native projects alongside rendered media. Tasks span architecture, characters, products, animation and more, with requirements on both appearance and the structure behind it.
78 Text
130 Image
15 Video
59 tasks also include specification attachments. All agents are evaluated on the same 223 tasks.
02 The results
How well do agents create?
The complete main table, with scores on a 0–100 scale.
Highest overall score79.9/100
GPT-6-Astra leads the benchmark.
Fully satisfied rubrics42.2%
Even the leading model passes every criterion on fewer than half of all tasks.
Multimodal advantage+8.6pts
GPT-6-Astra over Opus-5 on image and video inputs.
How to read this table
Overall Mean task score across all 223 tasks; each task combines 50% Rubric and 50% Dimension.
Text / MM Mean scores on 78 text tasks and 145 multimodal tasks (130 image + 15 video).
Rubric / Dimension Weighted binary requirement satisfaction / mean quality across eight visual dimensions, each normalized to 0–100.
Full pass Percentage of tasks that pass every rubric criterion. It does not mean an overall score of 100.
Turns / Tools Mean agent turns and tool calls per task, as reported in the manuscript.
Output tokens Mean generated tokens per task, in thousands (K). Failed evaluations receive zero and remain in score denominators.
Look beyond the average
Different worlds,
different strengths.
Select a category to compare all 13 models.
Mean overall score / 100
Category scores use all tasks in the selected category, including failed evaluations.
03 The artifacts
A closer look at the work.
Open a scene. Inspect the brief. Turn it around.
Explore 181 submissions from seven models across 26 tasks, spanning expansive environments, animations and focused objects. Choose a model to explore its works, or open a task to compare up to seven submissions, with interactive 3D previews and paper-matched scores.
These are curated examples, not a random sample or a measure of average quality. Scores are from the frozen paper runs; 3D previews are derived from those same projects.
04 The evaluation
More than a good-looking render.
Check what a scene does, as well as how it looks.
Complementary evidence
An independent rendering agent selects eight task-relevant views. Structure is checked through Blender tools; appearance is judged using renders and video frames.
Requirements + quality
Task-specific binary criteria measure requirement satisfaction. Eight visual dimensions measure quality. GPT-5.6-terra serves as the main automatic judge.
The task score
Overall = 0.5 × Rubric + 0.5 × Dimension
Both components on a 0–100 scaleSee the evaluation pipeline
A check against human preferences
82.99% agreement.
The main judge’s overall scores agree with the majority human preference on 122 of 147 eligible pairs.
150 same-input pairs were rated by three reviewers. Skips count as ties; three pairs without a majority are excluded. A judge agrees only when its A-win, B-win or tie prediction matches the majority exactly.
| Judge model | Agreement ↑ | Spearman ρ ↑ |
|---|
Spearman correlation compares score differences sA − sB with majority labels A = +1, tie = 0, B = −1, over the same 147 pairs. Full Pearson and Kendall results are in the paper.