UStyle: Waterbody Style Transfer of Underwater Scenes by Depth-Guided Feature Synthesis

Md Abu Bakr Siddique, Vaishnav Ramesh, Junliang Liu, Piyush Singh, Md Jahidul Islam

---

Pointers

• Preprint: https://arxiv.org/pdf/2503.11893
• UF7D Dataset: DropBox Link
• UStyle checkpoints and qualitative comparison files are also in the same folder

Models and Files

1. The model is defined in model.py
2. Download and save the checkpoints in the checkpoints/ directory.
3. Our fusion model integrates content and style features via a depth-aware whitening and coloring transform (DA-WCT) blending. This model enhances waterbody stylization by fusing features from multiple scales while incorporating depth information
   • The fusion models are implemented in fusion_1to1.py and fusion_all.py.
   • Guided filtering for post-processing is implemented in utils/photo_gif.py (adapted from the PhotoWCT code).
4. train.py is the training code for our ResNet-based blockwise model.
5. Fine-tuning can be performed using finetune.py.

---

Requirements

• Python 3.10
• PyTorch (tested with version torch 1.8.1+cu111)
• torchvision 0.9.1+cu111, numpy, opencv-python, Pillow, and other standard libraries

---

Usage

1. Clone the repository:

   git clone https://github.com/uf-robopi/UStyle.git
   cd UStyle/

2. Setup the environment:

   conda env create -f environment.yml
   conda activate UStyle

3. Train and Finetune UStyle:

   python3 train.py
   python3 finetune.py

4. Inference using UStyle:

   python3 fusion_1to1.py
   python3 fusion_all.py
   

---

Bibliography

@article{siddique2025ustyle,
    author={Siddique, Md Abu Bakr and Ramesh, Vaishnav and Liu, Junliang and Singh, Piyush and Islam, Md Jahidul},
    title={UStyle: Waterbody Style Transfer of Underwater Scenes by Depth-Guided Feature Synthesis},
    journal={Accepted for publication in the IEEE Journal of Oceanic Engineering (JOE)},
    year={2025}
}
  

---

Acknowledgements

• https://github.com/NVIDIA/FastPhotoStyle
• https://github.com/chiutaiyin/PhotoWCT2