README.md

June 9, 2025 · View on GitHub

Q-Eval-100K: Evaluating Visual Quality and Alignment Level for Text-to-Vision Content

Zicheng Zhang¹, Tengchuan Kou¹, Shushi Wang¹, Chunyi Li¹, Wei Sun¹, Wei Wang²

Xiaoyu Li², Zongyu Wang², Xuezhi Cao², Xiongkuo Min¹, Xiaohong Liu¹, Guangtao Zhai¹

¹Shanghai Jiao Tong University, ²Meituan

Paper Available at Arxiv.

Dataset is now available at HF. Only the training instances are open-sourced.

The retrained Q-Eval-Score weight on the training instances is at https://huggingface.co/AGI-Eval-Official/Q-Eval-Score

Q-Eval serves as the dataset for the NTIRE 2025 XGC Track 2.

Motivation

Evaluating text-to-vision content hinges on two crucial aspects: visual quality and alignment. While significant progress has been made in developing objective models to assess these dimensions, the performance of such models heavily relies on the scale and quality of human annotations. According to Scaling Law, increasing the number of human-labeled instances follows a predictable pattern that enhances the performance of evaluation models. Therefore, we introduce a comprehensive dataset designed to Evaluate Visual quality and Alignment Level for text-tovision content (Q-EVAL-100K), featuring the largest collection of human-labeled Mean Opinion Scores (MOS) for the mentioned two aspects. The Q-EVAL-100K dataset encompasses both text-to-image and text-to-video models, with 960K human annotations specifically focused on visual quality and alignment for 100K instances (60K images and 40K videos). Leveraging this dataset with context prompt, we propose Q-Eval-Score, a unified model capable of evaluating both visual quality and alignment with special improvements for handling long-text prompt alignment. Experimental results indicate that the proposed Q-Eval-Score achieves superior performance on both visual quality and alignment, with strong generalization capabilities across other benchmarks. These findings highlight the significant value of the Q-EVAL-100K dataset.

Visual Quality Performance

Text Alignment Performance

Dataset Access

The training dataset is available at HF.

Model Release

Due to the Meituan copyright policies, currently we are not allowed to release the Q-Eval-Score model.

However, to support the research community, we have re-trained the model from scratch using the public part of Q-Eval data only. The weights of this fully open-source version are now available on Hugging Face:

👉 https://huggingface.co/AGI-Eval-Official/Q-Eval-Score

We hope this model can serve as a starting point for building strong and explainable visual evaluators.

💻 Inference

We provide a Python script infer.py for running inference using the open-source Q-Eval-Score model.

Task	PLCC	SRCC
Image Alignment	0.797	0.826
Image Quality	0.760	0.747
Video Alignment	0.613	0.614
Video Quality	0.700	0.673

The current performance is obtained by retesting our open-sourced version of the Q-Eval-Score model, which is trained entirely on publicly available data. This version does not include any proprietary annotations or Meituan internal data, which were used in the original Q-Eval release.

Citation

If you find our work useful, please cite our paper as:

@misc{zhang2025qeval100kevaluatingvisualquality,
      title={Q-Eval-100K: Evaluating Visual Quality and Alignment Level for Text-to-Vision Content}, 
      author={Zicheng Zhang and Tengchuan Kou and Shushi Wang and Chunyi Li and Wei Sun and Wei Wang and Xiaoyu Li and Zongyu Wang and Xuezhi Cao and Xiongkuo Min and Xiaohong Liu and Guangtao Zhai},
      year={2025},
      eprint={2503.02357},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2503.02357}, 
}