Seeing Through the Rain: Resolving High-Frequency Conflicts in Deraining and Super-Resolution via Diffusion Guidance
Wenjie Li Jinglei Shi, Jin Han, Heng Guo, Zhanyu Ma · 2025
Read Seeing Through the Rain: Resolving High-Frequency Conflicts in Deraining and Super-Resolution via Diffusion Guidance by Wenjie Li Jinglei Shi, Jin Han, Heng Guo, Zhanyu Ma (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-12419.txt.
Introduction
Wenjie Li and four co-authors tackle two image tasks that pull in opposite directions: removing rain and raising resolution. Rain removal deletes high-frequency detail, while super-resolution tries to rebuild it, so chaining the two often gives inconsistent results. Clean, sharp images matter for tasks such as spotting small objects.
Their model, DHGM, uses guidance from a pretrained diffusion model together with high-pass filters that isolate fine detail. That pairing lets it strip rain artifacts and enhance structure at the same time, rather than in conflicting stages. Taking deraining as the test case, experiments show better performance than existing methods, at lower cost, in the authors' evaluation.
Catalog record
| CREATOR | Wenjie Li Jinglei Shi, Jin Han, Heng Guo, Zhanyu Ma |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 38 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2511-12419
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Seeing Through the Rain: Resolving High-Frequency Conflicts in Deraining and Super-Resolution via Diffusion Guidance Wenjie Li Jinglei Shi, Jin Han, Heng Guo, Zhanyu Ma Abstract Clean images are crucial for visual tasks such as small object detection, especially at high resolutions. However, real-world images are often degraded by adverse weather, and weather restoration methods may sacrifice high-frequency details critical for analyzing small objects. A natural solution is to apply super-resolution (SR) after weather removal to recover both clarity and fine structures. However, simply cascading restoration and SR struggle to bridge their inherent conflict: removal aims to remove high-frequency weather-induced noise, while SR aims to hallucinate high-frequency textures from existing details, leading to inconsistent restoration contents. In this paper, we take deraining as a case study and propose DHGM, a Diffusion-based High-frequency Guided Model for generating clean and high-resolution images. DHGM integrates pre-trained diffusion priors with high-pass filters to simultaneously remove rain artifacts and enhance structural details. Extensive experiments demonstrate that DHGM achieves superior performance over existing methods, with lower costs. Code link: https://github.com/PRIS-CV/DHGM . Introduction Adverse weather commonly degrades visual quality, significantly impacting downstream tasks such as object detection ( 42 ) . To mitigate these degradations, weather restoration is typically employed to remove weather-induced noise and recover clean images. However, as pointed out in recent studies ( 48 ; 14 ) , existing methods inevitably sacrifice high-frequency details, resulting in excessive smoothing. This observation is also confirmed by our spectral visualization at the top of Fig. 1 , showing that high-frequency parts in weather removal images are attenuated compared to ground truth (GT). This indicates that existing methods remove not only weather noise but also high-frequency textures. Such an issue severely impairs detection of small targets, such as distant vehicles, since even at 2K resolutions, these objects are low-resolution, typically occupy only a few dozen pixels, making accurate detection particularly sensitive to texture or edge loss introduced by weather removal. \begin{overpic}[img/teaser_small.pdf] \end{overpic} Figure 1: ( Top ) Frequency-domain analysis shows that weather removal methods eliminate not only rain streaks but also valuable high-frequency textures. ( Bottom ) Compared with existing methods, our method better preserves details and improves small-object detection under rainy conditions. To preserve and reconstruct essential high-frequency textures for reliable downstream detection, an intuitive solution is to apply super-resolution (SR) ( 56 ) after weather removal. Unlike general image restoration methods, which primarily focus on noise removal or texture enhancement without explicitly upsampling images, SR methods directly upscale image resolution, thus effectively enlarging small targets to reduce hallucination, as shown in Fig. 1 . However, simply cascading weather removal and SR methods fails to address their inherent conflicts: weather removal aims to suppress high-frequency noise, whereas SR attempts to infer high-frequency details from existing textures. Errors introduced in weather removal propagate through SR, resulting in amplified artifacts and inconsistent texture reconstruction. Similarly, fine-tuning deraining models ( 38 ) on paired low-resolution (LR) rainy and high-resolution (HR) clean images faces similar issues, as these models inherently struggle to balance high-frequency noise removal and texture recovery. Consequently, there remains a critical need for a method that can simultaneously achieve effective weather noise suppression and accurate SR texture restoration. In this paper, we focus on image deraining as a representative example to address this challenging scenario. \begin{overpic}[img/teaser2_new_new.pdf] \par\put(5.0,0.5){\color[rgb]{0,0,0}{(a) Synthetic Rain Test set}} \put(25.7,0.5){\color[rgb]{0,0,0}{(b) Synthetic Rain+Raindrop Test set}} \put(55.5,0.5){\color[rgb]{0,0,0}{(c) Real Raindrop Test set}} \put(82.5,0.5){\color[rgb]{0,0,0}{(d) Real Rain Test set}} \par\put(0.0,19.99){\color[rgb]{0,0,0}{(a) Our insight: guided and high-pass filter for high-frequency textures preserving while removing rain effects.}} \put(76.5,19.99){\color[rgb]{0,0,0}{(b) Pipeline of existing Methods.}} \par\put(-2.0,24.1){\color[rgb]{0,0,0}{{\rotatebox{90.0}{{Frequency Analysis}}}}} \put(-2.0,4.8){\color[rgb]{0,0,0}{{\rotatebox{90.0}{{Efficiency Trad-off}}}}} \par\par\end{overpic} Figure 2: ( Top ) Given a clean and HR image, guided filters can remove high-frequency noises while preserving most high-frequency textures, and high-pass filters further enhance blurred high-frequency edges. ( Bottom ) Our method achieves the best performance while requiring less cost in terms of speed and model size. (* denotes results of fine-tuning on our training dataset.) Inspired by guided filters ( 11 ) and high-pass filters ( 16 ) , we try to utilize these to reconstruct rainy LR images. Specifically, as shown in Fig. 2 , with the help of clean and HR images, we observe guided filter ( 11 ) can smooth out messy high-frequency noise ( e . g ., rain and raindrop) while preserving image edge ( e . g ., high-frequency textures of LR images), thereby aligning the frequency distribution is close to the ground truth (GT). However, high-frequency texture details from restored images remain missing, as shown in the spatial results and frequency distribution. Therefore, a high-pass filter can be further applied to clean HR images to compensate for high-frequency texture features, leading to rain-free and HR outputs. Compared to existing deraining and SR methods, this strategy achieves cleaner and clearer results. Therefore, combining priors with guided and high-pass filters can be a promising solution for SR in rainy weather to address the balance in high-frequency reconstruction. However, clean and HR images are required by guided and high-pass filters. To achieve this chicken-egg problem, we try to learn content priors close to GT distributions. Specifically, we propose a Diffusion-based High-frequency Guided Model (DHGM) with two phases. In the first phase, we employ encoders to compress contents reflecting true distributions into latent spaces as priors. To leverage latent priors for rain removal and texture reconstruction, we propose a Media Remover (MR) based on guided filters and a Texture Compensator (TC) based on high-pass filters. Recognizing the potential of diffusion models ( 12 ) (DM) to achieve high-quality mappings from randomly sampled Gaussian noise to latent distributions ( 36 ; 46 ) , we proceed to the second phase by freezing encoder weights and using DM to learn content distributions within pre-trained latent priors. Simultaneously, we fine-tune our MR and TC from the first phase, training them alongside DM to reconstruct images jointly. As observed at the bottom of Fig. 1 and Fig. 2 , our method can reconstruct clean and HR images from rainy LR images while maintaining efficiency on speed and model size, further improving the accuracy of downstream tasks. To summarize, our contributions are as follows: • We focus on deraining as an example, propose DHGM that explores the challenge of recovering clean and HR images from potential LR images captured in rainy conditions; • We propose MR and TC modules based on guided filter and high-pass filter, which direct pre-trained diffusion priors to remove rain-induced noise and recover textures; • Experiments on extensive datasets show our method can recover clean and HR images for downstream tasks while requiring less computational cost than existing methods. Related Work \begin{overpic}[img/main.pdf] \end{overpic} Figure 3: Overview of our method, which utilizes our Media Remover (MR) and Texture Compensator (TC) to guide learned diffusion priors in latent spaces to complete high-frequency rain-induced media removal and high-frequency texture reconstruction. Restoration in Rainy Weather Image Deraining. To handle rainy images that obstruct the view and are not conducive to downstream tasks ( 31 ) , a series of specially designed networks ( 33 ) employ strategies like multi-branch ( 13 ) , multi-scale ( 2 ) , and multi-stage ( 44 ) to achieve end-to-end deraining. Uddin et al . ( 40 ) focuses on