TextEraseBench

Paper | Code

TextEraseBench is a paired benchmark for scene-text removal. It contains 185 samples drawn from a mixture of synthetic scenes and real-world photographs captured by the authors.

Construction

1. Fine-grained text regions were manually annotated in each source image.
2. Nano Banana 2 was used to remove the selected text and generate the paired text-free background.
3. The resulting pairs were checked again to remove failed edits and visible artifacts.

Data Structure

TextEraseBench/
├── shot/    # Source images containing text
├── bg/      # Paired text-free backgrounds
└── mask/    # Fine-grained text masks

Files with the same relative name form one evaluation sample.

Evaluation

Use shot/ and mask/ as model inputs and bg/ as the reconstruction target. Report results over the complete 185-sample set and keep preprocessing settings consistent across compared methods.

License

TextEraseBench is released under CC BY 4.0 to the extent that the dataset authors hold rights in the images and annotations. Some paired backgrounds were produced using the Gemini/Nano Banana service; users remain responsible for complying with any applicable service terms and for assessing their intended use.

Citation

@inproceedings{zhou2026osor,
  title     = {OSOR: One-Step Diffusion Inpainting for Effect-Aware Object Removal},
  author    = {Zhou, Qinming and Sun, Chenxi and Kong, Deyang and He, Junhao and Tang, Xiangheng and Yu, Peike and Wu, Haotian and Cao, Leilei and Zhang, Linfeng},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
  url       = {https://arxiv.org/abs/2606.28094}
}