Low-light Image Enhancement via CLIP-Fourier Guided Wavelet Diffusion

Minglong Xue, Jinhong He, Yanyi He, Zhipu Liu, Wenhai Wang, Mingliang Zhou

Abstract

Low-light image enhancement techniques have significantly progressed, but unstable image quality recovery and unsatisfactory visual perception are still significant challenges. To solve these problems, we propose a novel and robust low-light image enhancement method via CLIP-Fourier Guided Wavelet Diffusion, abbreviated as CFWD. Specifically, CFWD leverages multimodal visual-language information in the frequency domain space created by multiple wavelet transforms to guide the enhancement process. Multi-scale supervision across different modalities facilitates the alignment of image features with semantic features during the wavelet diffusion process, effectively bridging the gap between degraded and normal domains. Moreover, to further promote the effective recovery of the image details, we combine the Fourier transform based on the wavelet transform and construct a Hybrid High Frequency Perception Module (HFPM) with a significant perception of the detailed features. This module avoids the diversity confusion of the wavelet diffusion process by guiding the fine-grained structure recovery of the enhancement results to achieve favourable metric and perceptually oriented enhancement. Extensive quantitative and qualitative experiments on publicly available real-world benchmarks show that our approach outperforms existing state-of-the-art methods, achieving significant progress in image quality and noise suppression.

Introduction

Low-light image enhancement (LLIE) is a classic computer vision challenge: recovering color, contrast, and detail from underexposed images without introducing noise or artifacts. While deep learning has made strides, many models still struggle with color distortion or lose fine-grained textures in shadows. Traditional methods often rely on manually designed priors that fail to adapt to diverse real-world lighting, while modern CNNs and Transformers often supervise only at the image level, leading to "diversity confusion" where the model fails to reconstruct local structures accurately.

Diffusion Models offer high-quality generation but are computationally expensive and can be "too creative," adding redundant information. The proposed CFWD framework addresses these issues by combining the generative power of diffusion with multimodal semantic guidance (CLIP) and frequency-domain constraints (Fourier and Wavelets).

Methodology

The CFWD framework stands on three main pillars designed to ensure structural accuracy and visual naturalness:

  1. Wavelet Diffusion Model: Instead of processing high-resolution pixels, the model applies a Discrete Wavelet Transform (DWT). It performs the noisy-to-clean iteration only on the low-frequency subband, cutting spatial dimensions to 1/4 and significantly reducing computational load while preserving core content.
  2. Multiscale Visual-Language Guidance (CLIP): The model uses a pre-trained CLIP model to align image features with text prompts. Positive prompts (e.g., "clear, bright") and negative prompts (e.g., "dark, noisy") act as semantic anchors, pulling the enhancement toward high quality and pushing it away from degradation artifacts.
  3. Hybrid High Frequency Perception Module (HFPM): To ensure fine details are not lost, the authors combine Wavelet and Fourier transforms. By minimizing differences in amplitude and phase of high-frequency components between the enhanced image and ground truth, the model achieves sharper structural recovery.
The overall architecture of the proposed CFWD framework, illustrating the wavelet diffusion process and multimodal guidance.

Experimental Results

CFWD was rigorously tested on paired datasets (LOLv1, LOLv2-Real, LSRW), high-resolution benchmarks (BAID), and unpaired sets (LIME, DICM). The model consistently outperformed state-of-the-art (SOTA) methods across both quantitative and qualitative metrics.

  • Quantitative Performance: On the LOLv1 dataset, CFWD achieved a PSNR of 29.185 dB, a significant jump over previous leaders like GSAD (27.629 dB). It also scored the lowest (best) on LPIPS and FID metrics, indicating results that appear more natural to human observers.
  • High-Resolution Robustness: On 2K resolution images (BAID), it maintained a PSNR of 26.918 dB, proving its practicality for modern photography standards.
  • Visual Quality: Qualitative comparisons demonstrate that while other models often overexpose areas or leave "color clouds" in shadows, CFWD produces clean, balanced, and sharp results.
Qualitative comparison of CFWD against state-of-the-art methods on various low-light scenes, highlighting superior detail recovery and noise suppression.

Conclusion

By shifting the heavy lifting of diffusion to the wavelet domain and utilizing semantic anchors via CLIP, CFWD effectively bridges the gap between degraded and normal domains. While diffusion models remain slower than simple CNN-based enhancers, the use of frequency-domain synergy provides a powerful constraint against noise and hallucination. Future work will focus on automated prompt engineering and further reducing computational overhead to make these advanced generative techniques even more accessible for real-time applications.

Summary updated

Abstract

Low-light image enhancement techniques have significantly progressed, but unstable image quality recovery and unsatisfactory visual perception are still significant challenges. To solve these problems, we propose a novel and robust low-light image enhancement method via CLIP-Fourier Guided Wavelet Diffusion, abbreviated as CFWD. Specifically, CFWD leverages multimodal visual-language information in the frequency domain space created by multiple wavelet transforms to guide the enhancement process. Multi-scale supervision across different modalities facilitates the alignment of image features with semantic features during the wavelet diffusion process, effectively bridging the gap between degraded and normal domains. Moreover, to further promote the effective recovery of the image details, we combine the Fourier transform based on the wavelet transform and construct a Hybrid High Frequency Perception Module (HFPM) with a significant perception of the detailed features. This module avoids the diversity confusion of the wavelet diffusion process by guiding the fine-grained structure recovery of the enhancement results to achieve favourable metric and perceptually oriented enhancement. Extensive quantitative and qualitative experiments on publicly available real-world benchmarks show that our approach outperforms existing state-of-the-art methods, achieving significant progress in image quality and noise suppression. The project code is available at https://github.com/hejh8/CFWD.
BibTeX
@article{xue2025low,
  author = {Minglong Xue and Jinhong He and Yanyi He and Zhipu Liu and Wenhai Wang and Mingliang Zhou},
  title = {{Low-light Image Enhancement via CLIP-Fourier Guided Wavelet Diffusion}},
  journal = {ACM Trans. Multim. Comput. Commun. Appl.},
  year = {2025},
  doi = {10.1145/3764933},
}
Figure