Home

Awesome

CRS-Diff: Controllable Generative Remote Sensing Foundation Model

Paper (ArXiv)

<div align=center> <img src="img/figure_1.png" height="100%" width="100%"/> </div>

TODO

Environment

conda env create -f environment.yaml
conda activate csrldm

You can download pre-trained models last.ckpt and put it to ./ckpt/ folder.

Testing

You can run the code to start the gradio interface by:

python src/test/test.py

The demonstration effects of the project are as follows:

<div align=center> <img src="img/figure_2.png" height="100%" width="100%"/> </div>

You can also use the following code to generate images more quickly

python src/test/inference.py

Some of the results are shown below：

<div align=center> <img src="img/figure_3.png" height="100%" width="100%"/> </div>

Acknowledgments:

This repo is built upon ControlNet and Uni-ControlNet. Some of the functional implementations of remote sensing imagery refer to: GeoSeg,Txt2Img-MHN and SGCN. Sincere thanks to their excellent work!

Citation

@misc{tang2024crsdiff,
      title={CRS-Diff: Controllable Generative Remote Sensing Foundation Model}, 
      author={Datao Tang and Xiangyong Cao and Xingsong Hou and Zhongyuan Jiang and Deyu Meng},
      year={2024},
      eprint={2403.11614},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}