[IJCNN 2025] ResQu: Quaternion Wavelet-Conditioned Diffusion Models for Image Super-Resolution

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

Luigi Sigillo, Christian Bianchi, Aurelio Uncini, and Danilo Comminiello

ISPAMM Lab, Sapienza University of Rome

📰 News

• [2025.07.05] Presented the work at IJCNN 2025 in Rome!
• [2025.06.05] Checkpoints and code are released!
• [2025.05.05] The paper has been published on Arxiv 🎉. The pdf version is available here!
• [2025.03.31] The paper has been accepted for presentation at IJCNN 2025 🎉!

😮 Highlights

💡 Elevating Image Super-Resolution with Quaternion Wavelets

Our work introduces ResQu, a novel approach that significantly advances image super-resolution by leveraging the power of quaternion wavelet embeddings. This allows for superior feature representation, leading to high-fidelity reconstructions and enhanced perceptual quality, a crucial step in various computer vision applications.

🔥 State-of-the-Art Performance with Novel Conditioning

We propose a streamlined framework that conditions a latent diffusion model (built upon the StableSR baseline) using quaternion wavelet embeddings. Through extensive experimentation, ResQu demonstrates a +15% PSNR improvement over traditional super-resolution models, showcasing its state-of-the-art capabilities in capturing intricate texture details.

👀 A Multi-Scale and Frequency-Aware Approach

Unlike existing methods that demand heavy preprocessing, complex architectures, and additional components like captioning models, our approach is efficient and straightforward. This enables a new frontier in real-time BCIs, advancing tasks like visual cue decoding and future neuroimaging applications.

🚀 Main Results

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

How to run experiments 💻

Building Environment

conda create --name=resqu python=3.9
conda activate resqu

pip install torch torchvision torchaudio --index-url [https://download.pytorch.org/whl/cu118](https://download.pytorch.org/whl/cu118)
pip install diffusers transformers accelerate xformers==0.0.16 wandb numpy==1.26.4 datasets scikit-learn torchmetrics==1.4.1 scikit-image pytorch_fid

Train

To launch the training of the model, you can use the following command, you need to change the output_dir and also specify the gpu number you want to use, right now only 1 GPU is supported:

CUDA_VISIBLE_DEVICES=N accelerate launch src/resqu/train_resqu.py \
    --pretrained_model_name_or_path=stabilityai/stable-diffusion-2-1-base \
    --output_dir=output/resqu_model_out \
    --dataset_name=your_huggingface_dataset_name \
    --image_column=image \
    --conditioning_column=quaternion_wavelet_embedding \
    --resolution=512 \
    --learning_rate=1e-5 \
    --train_batch_size=8 \
    --num_train_epochs=50 \
    --tracker_project_name=resqu \
    --enable_xformers_memory_efficient_attention \
    --checkpointing_steps=1000 \
    --validation_steps=500 \
    --report_to wandb

Generate

Request access to the pretrained models from Google Drive.

To launch the generation of the images from the model, you can use the following commands:

CUDA_VISIBLE_DEVICES=N python src/resqu/generate_resqu.py \
    --model_path=output/resqu_model_out/checkpoint-XXXXX/ \
    --input_low_res_image_path=path/to/your/low_res_image.png \
    --output_dir=generated_images/

Evaluation

Request access to the pretrained models from Google Drive.

To launch the testing of the model, you can use the following command, you need to change the output_dir:

CUDA_VISIBLE_DEVICES=N python src/resqu/evaluation/evaluate.py \
    --generated_images_path=generated_images/ \
    --ground_truth_images_path=path/to/your/ground_truth_images/

Cite

Please cite our work if you found it useful:

@misc{sigillo2025quaternionwaveletconditioneddiffusionmodels,
      title={Quaternion Wavelet-Conditioned Diffusion Models for Image Super-Resolution}, 
      author={Luigi Sigillo and Christian Bianchi and Aurelio Uncini and Danilo Comminiello},
      year={2025},
      eprint={2505.00334},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2505.00334}, 
}

Star History

Acknowledgement

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