Controllable UWFA Editing

This is the repository for "Controllable editing via diffusion inversion on ultra-widefield fluorescein angiography for the comprehensive analysis of diabetic retinopathy".

CLIP Tuning

In the training stage, the CLIP is tuned by prompt tuning strategy using infoNCE loss.

sh scripts/train_clip.sh

Stable Diffusion Tuning

In the training stage, the SD model is trained with the multimodal embeddings.

sh scripts/multi_lora.sh

Controllable Editing

During the inferencing phase, the original UWFA image is edited into disease-free domain.

sh scripts/inference_inversion.sh

Here is an example, which includes the original image, the edited image, and the difference between the two.

Citing

If this project is help for you, please cite it.

@article{ma2024controllable,
  title={Controllable editing via diffusion inversion on ultra-widefield fluorescein angiography for the comprehensive analysis of diabetic retinopathy},
  author={Ma, Xiao and Ji, Zexuan and Chen, Qiang and Ge, Lexin and Wang, Xiaoling and Chen, Changzheng and Fan, Wen},
  journal={Biomedical Optics Express},
  volume={15},
  number={3},
  pages={1831--1846},
  year={2024},
  publisher={Optica Publishing Group}
}