Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quaternion Wavelet-Conditioned Diffusion Models for Image Super-Resolution

Luigi Sigillo, Christian Bianchi, Aurelio Uncini, Danilo Comminiello

TL;DR

This work tackles image super-resolution by balancing perceptual quality with structural fidelity in high upscaling scenarios. It introduces ResQu, a latent-diffusion SR framework that injects quaternion wavelet embeddings via a time-aware encoder into a frozen Stable Diffusion backbone, enabling multi-scale conditioning at denoising time steps. Key contributions include a novel quaternion wavelet and time-aware encoder, a conditioning mechanism that fuses QUAVE features with timestep information, and extensive ablations plus cross-domain demonstrations showing robust improvements over state-of-the-art methods. The approach leverages generative priors from foundation models to achieve high-fidelity, texture-rich reconstructions, with strong performance on RealSR, DRealSR, and DIV2K, plus zero-shot generalization to ShipSpotting data.

Abstract

Image Super-Resolution is a fundamental problem in computer vision with broad applications spacing from medical imaging to satellite analysis. The ability to reconstruct high-resolution images from low-resolution inputs is crucial for enhancing downstream tasks such as object detection and segmentation. While deep learning has significantly advanced SR, achieving high-quality reconstructions with fine-grained details and realistic textures remains challenging, particularly at high upscaling factors. Recent approaches leveraging diffusion models have demonstrated promising results, yet they often struggle to balance perceptual quality with structural fidelity. In this work, we introduce ResQu a novel SR framework that integrates a quaternion wavelet preprocessing framework with latent diffusion models, incorporating a new quaternion wavelet- and time-aware encoder. Unlike prior methods that simply apply wavelet transforms within diffusion models, our approach enhances the conditioning process by exploiting quaternion wavelet embeddings, which are dynamically integrated at different stages of denoising. Furthermore, we also leverage the generative priors of foundation models such as Stable Diffusion. Extensive experiments on domain-specific datasets demonstrate that our method achieves outstanding SR results, outperforming in many cases existing approaches in perceptual quality and standard evaluation metrics. The code will be available after the revision process.

Quaternion Wavelet-Conditioned Diffusion Models for Image Super-Resolution

TL;DR

This work tackles image super-resolution by balancing perceptual quality with structural fidelity in high upscaling scenarios. It introduces ResQu, a latent-diffusion SR framework that injects quaternion wavelet embeddings via a time-aware encoder into a frozen Stable Diffusion backbone, enabling multi-scale conditioning at denoising time steps. Key contributions include a novel quaternion wavelet and time-aware encoder, a conditioning mechanism that fuses QUAVE features with timestep information, and extensive ablations plus cross-domain demonstrations showing robust improvements over state-of-the-art methods. The approach leverages generative priors from foundation models to achieve high-fidelity, texture-rich reconstructions, with strong performance on RealSR, DRealSR, and DIV2K, plus zero-shot generalization to ShipSpotting data.

Abstract

Image Super-Resolution is a fundamental problem in computer vision with broad applications spacing from medical imaging to satellite analysis. The ability to reconstruct high-resolution images from low-resolution inputs is crucial for enhancing downstream tasks such as object detection and segmentation. While deep learning has significantly advanced SR, achieving high-quality reconstructions with fine-grained details and realistic textures remains challenging, particularly at high upscaling factors. Recent approaches leveraging diffusion models have demonstrated promising results, yet they often struggle to balance perceptual quality with structural fidelity. In this work, we introduce ResQu a novel SR framework that integrates a quaternion wavelet preprocessing framework with latent diffusion models, incorporating a new quaternion wavelet- and time-aware encoder. Unlike prior methods that simply apply wavelet transforms within diffusion models, our approach enhances the conditioning process by exploiting quaternion wavelet embeddings, which are dynamically integrated at different stages of denoising. Furthermore, we also leverage the generative priors of foundation models such as Stable Diffusion. Extensive experiments on domain-specific datasets demonstrate that our method achieves outstanding SR results, outperforming in many cases existing approaches in perceptual quality and standard evaluation metrics. The code will be available after the revision process.
Paper Structure (11 sections, 2 equations, 5 figures, 3 tables)

This paper contains 11 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Background
  3. The Proposed ResQu Method
  4. Quaternion Wavelet and Time-Aware Encoder
  5. Latent Diffusion Process
  6. Experimental Results
  7. Experimental Setup
  8. Quantitative Evaluation
  9. Qualitative Analysis
  10. Ablation Studies
  11. Conclusion

Figures (5)

  • Figure 1: Qualitative comparisons on real-world images, upscaled from 128 to 512 pixels, we compare StableSR wang2023exploiting, and the ground truth with our results displayed in the final column. The visual evaluations underscore differences in image quality, clarity, and the enhancement of fine details achieved by our model during the super-resolution process. Noticeable to see that the background on StableSR results is plain while in the ground truth it is coarse, like in the output of our model.
  • Figure 2: Overview of ResQu Super-Resolution framework. We pre-trained QUAVE and trained only our encoder, the quaternion embeddings- and time- aware encoder. The other parts of the framework are frozen, which speeds up the overall training process.
  • Figure 3: Comparison of the LR input image with SR outputs generated by state-of-the-art methods and our proposed model on the DRealSR wei2020component and RealSR cai2019toward datasets. The red bounding box highlights a zoomed-in region, illustrating the superior resolution and detail preservation achieved by our method compared to existing approaches.
  • Figure 4: Impact of the number of sampling steps on key evaluation metrics. Reducing the steps improves efficiency while maintaining performance, highlighting the inherent speed advantage of our model.
  • Figure 5: Visual comparison of diverse super-resolution results generated by changing the CFW. The images demonstrate the impact of our CFW enhanced by our encoder, producing more structured details.