Frontier3DBench: Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds

Frontier3DBench evaluates coding agents on executable 3D creation, from individual objects and complete scenes to interactive open worlds. Agents generate editable Blender scenes or runnable Unreal Engine projects from text, images, or existing scenes, with reproducible code and required deliverables.

Benchmark overview

The release contains 145 tasks across three levels:

Level	Scope	Tasks	Runtime
L1	Object creation and reconstruction	65	Blender
L2	Scene creation, reconstruction and editing	65	Blender
L3	Interactive environments and open worlds	15	Unreal Engine

Evaluation covers geometry, delivery, visual fidelity, local detail and interaction. All three levels use the unified python -m benchmark interface. The overall score combines level means with 30% / 30% / 40% weights for L1 / L2 / L3. See the task guide and scoring guide for the evaluation contracts.

This dataset hosts the task packages from the GitHub repository, preserving data/3dbench/tasks and its asset attribution. Generation and evaluation code and scoring contracts are maintained in the GitHub repository. Use the installation guide to run the benchmark. The dataset snapshot and per-file checksums are recorded in data-manifest.json. Model submissions, experiment outputs, credentials and installed third-party runtimes are excluded. Candidate agents receive staged task requirements and inputs; organizer scoring details and reference-source attribution remain outside candidate workspaces.

Documentation

• Installation and running — source checkout, Git LFS, dependencies, model providers, Blender/UE setup, generation, evaluation and resume.
• Documentation index · 中文文档 — tasks, models, experiment settings, scoring and result exports.
• Agent instructions — workflow for an agent operating this benchmark.
• Security · Asset licenses — execution boundaries and task-specific attribution.

Citation

If you use Frontier3DBench, please cite:

@misc{lin2026frontier3dbench,
  title = {Frontier3DBench: Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds},
  author = {Lin, Xinjie and Liu, Qingle and Li, Wenyi and Lam, John and Wang, Xiaoqiu and Hong, Deyao and Gao, Mingju and Jiang, Zheng and Chi, Yizhe and Zheng, Youjie and Xiao, Calvin and Yang, Kaisen and Gao, Huan-ang},
  year = {2026},
  eprint = {2611.09017},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2611.09017}
}

License

Benchmark software is licensed under Apache-2.0. Task assets retain their individual licenses and attribution; see third-party notices and the asset license guide. Blender, model CLIs and Unreal Engine are installed separately.