WaterGen

WaterGen: Decoupling Scene and Medium in Underwater Image Generation

Jiayi Wu^1,, Tianfu Wang^1,, Tianyi Xiong¹, Dehao Yuan¹, Xiaomin Lin², Md Jahidul Islam³, Cornelia Fermüller¹, Christopher Metzler¹, Yiannis Aloimonos¹

¹University of Maryland  ·  ²University of South Florida  ·  ³University of Florida

^*Equal contribution

TL;DR

WaterGen generates diverse underwater images while independently controlling scene content and physical water-medium effects. A LoRA-adapted diffusion backbone first synthesizes a clean underwater scene latent; a medium-conditioned decoder then applies attenuation, scattering, and background light. The same scene can be rendered under many water conditions without changing geometry, producing paired clean/degraded images for downstream restoration and segmentation.

Overview

WaterGen treats underwater generation as two separate controls: what is in the image and what the water does to the image. Existing methods typically entangle these factors, so changing water color or turbidity often alters object layout. WaterGen avoids this by decoupling the two stages. A LoRA-adapted SDXL backbone generates a clean underwater scene latent from a text prompt, and a medium-conditioned decoder applies physically meaningful attenuation and backscattering according to specified water parameters.

This decoupled design lets a single generated scene be rendered under many water types without changing the underlying geometry. The resulting aligned clean/degraded pairs can be used as synthetic training data for underwater restoration and semantic segmentation.

Model Files

This repository contains the inference checkpoints used to reproduce the results shown on the project page.

stage1/
└── pytorch_lora_weights.safetensors   # SDXL LoRA for clean scene generation (rank 32, α 16)
stage2/
└── model.pth                          # medium-conditioned decoder

Refer to the code repository for loading and inference instructions.

Citation

@inproceedings{wu2026watergen,
  title={WaterGen: Decoupling Scene and Medium in Underwater Image Generation},
  author={Wu, Jiayi and Wang, Tianfu and Xiong, Tianyi and Yuan, Dehao and Lin, Xiaomin and Islam, Md Jahidul and Fermuller, Cornelia and Metzler, Christopher and Aloimonos, Yiannis},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2026}
}