BiliSakura/RSEdit-Benchmark-Results

RSCC-RSEdit-Test-Split

Project Page | Paper | GitHub

This repository contains the test split for RSEdit, a unified framework designed for text-guided image editing in the remote sensing (RS) domain.

Description

RSEdit adapts pre-trained text-to-image diffusion models (including U-Net and DiT architectures) into instruction-following editors for Earth observation imagery. This dataset consists of bi-temporal remote sensing image pairs and corresponding textual instructions. It is used to evaluate the model's ability to perform precise, physically coherent edits—such as simulating disaster impacts, urban growth, and seasonal shifts—while maintaining geospatial consistency and orthographic constraints.

The full framework was trained on over 60,000 semantically rich bi-temporal image pairs to learn these complex spatial and temporal priors.

Citation

If you find this dataset or the RSEdit framework useful, please cite the following work:

@misc{zhenyuan2026rsedittextguidedimageediting,
      title={RSEdit: Text-Guided Image Editing for Remote Sensing}, 
      author={Chen Zhenyuan and Zhang Zechuan and Zhang Feng},
      year={2026},
      eprint={2603.13708},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.13708}, 
}