Epiphany10086/GenBlemish-27K

GenBlemish-27K

GenBlemish-27K is a fine-grained dataset for detecting, localizing, and describing visual artifacts in text-to-image (T2I) generations. It contains 6,025 generated images and 27,507 annotated distortion regions covering 12 artifact categories. Each annotated region is paired with a pixel-level mask, a category label, and a natural-language description.

The dataset is designed for research on artifact perception, region-level diagnosis, visual-language grounding, and localized image correction.

Dataset Highlights

• 6,025 text-to-image images.
• 27,507 annotated distortion regions.
• 12 fine-grained artifact categories.
• Pixel-level region masks and natural-language descriptions.
• Input prompts paired with generated images.
• Images collected from outputs of more than 20 T2I systems, including Dreamina, Midjourney, Kandinsky, and SDXL.
• Human-AI collaborative annotation with professional revision and validation.

Intended Uses

GenBlemish-27K can be used for:

• Context-aware artifact localization.
• Fine-grained artifact classification.
• Vision-language model evaluation and grounding.
• Training or evaluating automated image-retouching and localized inpainting systems.
• Studying perceptual reliability of text-to-image generation.

Distortion Taxonomy

The dataset organizes artifacts into six high-level dimensions and twelve fine-grained categories.

The category names above follow the taxonomy described in the associated paper. The exact string values in the released annotation files should be treated as the source of truth.

Access Request

Access to the dataset requires completion of both of the following steps:

1. Complete the dataset access questionnaire.
2. Download and complete the NDA_Form, which is provided in the main branch of this repository.

Please send the completed NDA_norm.pdf to: shenshaocheng@sjtu.edu.cn. Dataset access will be provided after the request has been reviewed and approved.

Download

After your access request has been approved, the archive can be downloaded with the Hugging Face CLI:

hf download Epiphany10086/GenBlemish-27K GenBlemish.zip \
  --repo-type dataset \
  --local-dir ./GenBlemish-27K

Then extract the archive with a tool such as 7-Zip or WinRAR.

For Windows PowerShell, the equivalent command is:

hf download Epiphany10086/GenBlemish-27K GenBlemish.zip `
  --repo-type dataset `
  --local-dir .\GenBlemish-27K

Limitations and Responsible Use

• The dataset focuses on artifacts in AI-generated images and may not represent the full distribution of failures produced by all T2I systems.
• The category distribution is imbalanced; hand-related artifacts are particularly frequent.
• Annotations describe perceptual and semantic inconsistencies. They should not be interpreted as objective judgments about people, cultures, or real-world objects.
• Model names and generated content may be subject to third-party terms of use.
• Users are responsible for complying with the terms of use of the underlying image-generation systems and source materials.

License and Terms of Use

GenBlemish-27K is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license (CC BY-NC-SA 4.0).

When using, modifying, or redistributing the dataset, users must provide appropriate attribution, use the dataset for non-commercial purposes only, and distribute derivative works under the same license.

Users are also responsible for complying with the terms of use of the underlying image-generation systems and source materials.

Citation

If you use GenBlemish-27K, please cite the associated work:

@inproceedings{shen2026agenticretoucher,
  title     = {Agentic Retoucher for Text-to-Image Generation},
  author    = {Shen, Shaocheng and Liang, Jianfeng and Cai, Chunlei and Geng, Cong and Duan, Huiyu and Zhang, Xiaoyun and Hu, Qiang and Zhai, Guangtao},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}

Acknowledgements

We thank the annotators and reviewers who contributed to the human-AI collaborative construction and validation of GenBlemish-27K.

Contact

For questions, please open a discussion in this dataset repository or contact the corresponding authors of the associated paper: shenshaocheng@sjtu.edu.cn.