Text-to-3D Comprehensive Benchmark (T23D-CompBench) 🎥📊

Code · Project Page · Paper@ArXiv · Prompt list

Welcome to the T23D-CompBench dataset! This repository contains around 3,600 textured meshes generated by various models using the Prompt list. These textured meshes have been annotated from twelve evaluation dimensions, including Object Alignment, Attribute Alignment, Interaction Alignment, Overall Alignment, Texture Clarity, Texture Aesthetics, Geometry Loss, Geometry Redundancy, Geometry Roughness, Overall Visual Quality, 3D Authentic, and Overall Quality.

Dataset Details 📚

• Paper: Read the Paper

• Code: Code

• Prompt List (360 prompts): Prompt list

• Project Page: Project Page

• Models Included in T23D-CompBench dataset:

  • DreamFusion
  • Magic3D
  • 3DTopia
  • Consistent3D
  • One-2-3-45++
  • LGM
  • Hunyuan-2.0
  • Meshy_AI-5
  • Rodin_Gen-1.5
  • Tripo_AI-2.5

• Dataset Structure: Generate textured meshes are organized in the following structure (take 3DTopia for an example)

  ./3DTopia.zip/
  ├── A_asymmetrical_flower
    ├── model_normalized.obj
    ├── model_normalized.obj.mtl
    ├── material_0.png
  ├── A_bird_on_a_branch
    ├── model_normalized.obj
    ├── model_normalized.obj.mtl
    ├── material_0.png
  ...
  

Acknowledgements and Citation 🙏

This dataset is based on the text-to-3D generative framework, which utilizes various open-source repositories for textured mesh generation evaluation. If you find this dataset helpful, please consider citing the original work:

@article{cui2025towards,
title={Towards Fine-Grained Text-to-3D Quality Assessment: A Database and A Two-Stage Rank-Learning Metric},
author={Bingyang Cui, Yujie Zhang, Qi Yang, Zhu Li, and Yiling Xu},
journal={arXiv preprint arXiv:2509.23841},
year={2025}
}