[PRL 2026] Metadata, Wavelet, and Time Aware Diffusion Models for Satellite Image Super Resolution

If you like our project, please give us a star ⭐ on GitHub for latest update.

Luigi Sigillo, Renato Giamba, and Danilo Comminiello

ISPAMM Lab, Sapienza University of Rome

📰 News

• [2026.01.08] Accepted at Pattern Recognition Letters
• [2025.02.10] Accepted at ICLR 2025 Workshop on Machine Learning for Remote Sensing (ML4RS)

😮 Highlights

💡 Metadata, Wavelet, and Time-Aware Conditioning

MWT-Diff introduces a metadata-, wavelet-, and time-aware encoder that captures auxiliary attributes, multi-scale frequency cues, and temporal relationships to steer diffusion-based reconstruction for satellite image SR.

🔥 High-Fidelity Satellite Super-Resolution

By conditioning hierarchical diffusion dynamics with MWT-Encoder embeddings, the model restores fine-grained textures, boundaries, and spectral details that are critical for remote sensing analysis.

👀 Robust Across Datasets

Experiments show favorable performance over recent SR approaches on multiple satellite datasets, with improvements in perceptual quality metrics such as FID and LPIPS.

🚀 Main Results

For more evaluation, please refer to our paper for details.

How to run experiments 💻

Building Environment

conda create --name=mwt-diff python=3.9
conda activate mwt-diff

pip install -r requirements.txt

Train

To launch the training of the model, update the config paths and output directory as needed:

python main.py --base configs/sentinel/v2-finetune_text_T_512.yaml -t --gpus 1 --logdir logs/mwt_diff

Generate

Request access to the pretrained models from here.

To launch generation:

python scripts/sr_val_ddpm_text_T_vqganfin_old_sentinel.py \
	--init-img path/to/low_res_images \
	--outdir outputs/mwt_diff_samples \
	--config configs/sentinel/v2-finetune_text_T_512.yaml \
	--ckpt weights/stablesr_000117.ckpt \
	--vqgan_ckpt weights/vqgan_cfw_00011.ckpt \
	--ddpm_steps 200 \
	--colorfix_type adain

Evaluation

Request access to the pretrained models from here.

To run evaluation:

python -c "from basicsr.metrics.lpips_2dirs import compute_lpips_distance; compute_lpips_distance('outputs/mwt_diff_samples', 'path/to/ground_truth', output_file='lpips.txt', use_gpu=True)"

Cite

Please cite our work if you found it useful:

@article{SIGILLO2026102,
title = {Metadata, wavelet, and time aware diffusion models for satellite image super resolution},
journal = {Pattern Recognition Letters},
volume = {202},
pages = {102-107},
year = {2026},
issn = {0167-8655},
doi = {https://doi.org/10.1016/j.patrec.2026.01.031},
url = {https://www.sciencedirect.com/science/article/pii/S0167865526000589},
author = {Luigi Sigillo and Renato Giamba and Danilo Comminiello},
keywords = {Image super resolution, Diffusion models, Wavelet, Satellite image super resolution},
}

Star History

Acknowledgement

This project is based on StableSR. Thanks for their awesome work.