--- license: apache-2.0 task_categories: - image-to-3d - text-to-3d language: - en size_categories: - 1K

GEdit3D-Bench

A large-scale, open-world benchmark for versatile 3D asset editing.

GEdit3D-Bench construction and multi-dimensional evaluation overview.

**GEdit3D-Bench** is the evaluation benchmark for [Alchemy3D](https://github.com/libd1/Alchemy3D): an open-world 3D editing suite built from assets independent of the training data (newly synthesized content and curated Sketchfab assets). It contains **2,100** editing pairs over **6** edit types. Each item provides a **source asset**, **source / target captions**, **source / target images**, and an **editing instruction**. Evaluation jointly measures view quality, reference alignment, and MLLM-based scores. A ModelScope mirror is also available at [libd55/GEdit3D-Bench](https://modelscope.cn/datasets/libd55/GEdit3D-Bench). ## Dataset Layout The benchmark contains source assets (`glbs/`), condition images (`images/`), and metadata CSVs. `metadata.csv` indexes the **full** set (**2,100** pairs across `add`, `remove`, `replace`, `local_appearance`, `global_appearance`, `animation`). `metadata_small.csv` is the subset used for **MLLM / human scoring** (**420** pairs), kept smaller because these evaluations are expensive. Each row provides a source asset, source/target captions, source/target images, an editing instruction, and an edit type. ## Usage Download the benchmark: ```bash hf download libadi/GEdit3D-Bench --repo-type dataset --local-dir GEdit3D-Bench ``` Evaluation code and baselines would be released at **[edit3dstudio](https://github.com/libd1/edit3dstudio)**. ## License Apache License 2.0. Downstream 3D generators and renderers you apply to these assets may have their own terms. Sketchfab-sourced assets may carry additional terms from their original licenses. ## Citation ```bibtex @misc{li2026scalingversatile3dassets, title={Scaling Versatile 3D Assets Editing with a Million-Scale Dataset}, author={Badi Li and Tianxin Huang and Yu Zhou and Wei-Shi Zheng and Yi Ma and Shenghua Gao}, year={2026}, eprint={2609.34271}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2609.34271}, } ```