Flood-LDM: Generalizable Latent Diffusion Models for rapid and accurate zero-shot High-Resolution Flood Mapping

Accepted at WACV 2026: arXiv

Abstract

Flood prediction is critical for emergency planning and response to mitigate human and economic losses. Traditional physics-based hydrodynamic models generate high-resolution flood maps using numerical methods requiring fine-grid discretization; which are computationally intensive and impractical for real-time large-scale applications. While recent studies have applied convolutional neural networks for flood map super-resolution with good accuracy and speed, they suffer from limited generalizability to unseen areas. In this paper, we propose a novel approach that leverages latent diffusion models to perform super-resolution on coarse-grid flood maps, with the objective of achieving the accuracy of fine-grid flood maps while significantly reducing inference time. Experimental results demonstrate that latent diffusion models substantially decrease the computational time required to produce high-fidelity flood maps without compromising on accuracy, enabling their use in real-time flood risk management. Moreover, diffusion models exhibit superior generalizability across different physical locations, with transfer learning further accelerating adaptation to new geographic regions. Our approach also incorporates physics-informed inputs, addressing the common limitation of black-box behavior in machine learning, thereby enhancing interpretability.

Figure 1: High level overview of the process

Pretrained models

This section contains the latent diffusion model checkpoints. _gen.pth files contain the model parameters and are needed for inference, while _opt.pth files contain the saved optimizer and are needed for further transfer learning. The pretrained autoencoder checkpoint can be found here.

Training Catchment	CG-FG MSE	SR-FG-MSE	% change	Download
Wollombi	344.2	33.7	-90.19	checkpoint
Burnett	5957.4	731.2	-87.73	checkpoint
Chowilla	158.7	16.8	-89.40	checkpoint

After downloading the model checkpoints, edit the appropriate variables in the config file:

"resume_state": "/path/to/pretrained/model",
"vae_state": "/path/to/pretrained/vae",
"vae_dem_state": "/path/to/pretrained/dem/vae"

Do not include the _gen.pth/_opt.pth in the path to the pretrained LDM, the parser will automatically detect both files as long as they are in the same directory.

Requirements

Python 3.10.12+

pip install -r requirements.txt

Config

Different configuration options for training and testing of models are specified in a json file and can be passed into the training/testing command via a command-line argument. The config directory contains config files for reference, including examples of a standard diffusion model and a latent diffusion model.

2 diffusion model architectures are available in this codebase, adapted from the SR3 and DDPM papers. The desired model architecture can be selected using the which_model_G variable in the config file. It is recommended to use the sr3 architecture for standard diffusion models and the ddpm architecture for latent diffusion models.

Catchment data structure

Catchment data should be stored in this folder structure:

.
└── catchment_name/
    ├── train_hr/
    │   ├── Box_1_Depth_2021_03_18_00_00_00.tiff
    │   ├── Box_1_Depth_2021_03_18_00_30_00.tiff
    │   └── ...
    ├── train_lr/
    │   ├── Box_1_Depth_2021_03_18_00_00_00.tiff
    │   ├── Box_1_Depth_2021_03_18_00_30_00.tiff
    │   └── ...
    ├── test_hr/
    │   ├── Box_1_Depth_2015_04_21_00_00_00.tiff
    │   ├── Box_1_Depth_2015_04_21_00_30_00.tiff
    │   └── ...
    ├── test_lr/
    │   ├── Box_1_Depth_2015_04_21_00_00_00.tiff
    │   ├── Box_1_Depth_2015_04_21_00_30_00.tiff
    │   └── ...
    └── cropped_dems/
        ├── BOX_1_DEM.tiff
        ├── BOX_2_DEM.tiff
        └── ...

Training

Training is done with the following command:

python sr.py -p train -c config.json -output-dir path/to/output/dir

The model checkpoints will be stored in the output directory.

Testing

Testing is done with the following command:

python sr.py -p test -c config.json -output-dir path/to/output/dir

The super-resolution model outputs will be saved in the output directory if the corresponding option is set in the config file.

Citation

If you use any of the code provided in this repository or the models provided, please cite the following paper:

@misc{neo_2025_flood_ldm,
    title={Flood-LDM: Generalizable Latent Diffusion Models for rapid and accurate zero-shot High-Resolution Flood Mapping}, 
    author={Sun Han Neo and Sachith Seneviratne and Herath Mudiyanselage Viraj Vidura Herath and Abhishek Saha and Sanka Rasnayaka and Lucy Amanda Marshall},
    year={2025},
    month=nov,
    eprint={2511.14033},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2511.14033}, 
}

Acknowledgements

Code was adapted from this implementation of Image Super-Resolution via Iterative Refinement