💻 computer science

Single-Stage Signal Attenuation Diffusion Model for Low-Light Image Enhancement and Denoising

This paper proposes the Signal Attenuation Diffusion Model (SADM), a novel single-stage framework that integrates a signal attenuation mechanism into the diffusion process to simultaneously achieve brightness recovery and noise suppression for low-light image enhancement, thereby eliminating the need for the multi-stage pipelines or auxiliary networks required by existing methods.

Original authors: Ying Liu, Junchao Zhang, Caiyun Wu

Published 2026-04-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ying Liu, Junchao Zhang, Caiyun Wu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to look at a beautiful painting, but it's been covered in thick, dark fog, and someone has also sprinkled it with static electricity (noise). Your goal is to clean it up so you can see the colors and details again. This is what Low-Light Image Enhancement (LLIE) does for photos taken in the dark.

For a long time, computers tried to fix these photos in two separate steps:

Step 1: "Let's make it brighter!" (But this often made the static noise worse).
Step 2: "Okay, now let's clean up the noise." (But this often made the picture look blurry or washed out).

It's like trying to wash a muddy car by first spraying it with a hose (which spreads the mud everywhere) and then trying to dry it off. The two steps fight each other, and the result isn't perfect.

The New Idea: The "Signal Attenuation" Model (SADM)

The authors of this paper, Ying Liu and her team, came up with a smarter way. They built a Single-Stage Signal Attenuation Diffusion Model (SADM).

Here is how it works, using some creative analogies:

1. The "Reverse Movie" Analogy

Think of a Diffusion Model like a movie played in reverse.

The Forward Process (The Movie): Imagine taking a clear, bright photo and slowly turning the lights down until it's pitch black, while simultaneously adding static noise. Eventually, you have a screen full of random static.
The Reverse Process (The Fix): The AI watches this "movie" in reverse. It starts with the static noise and tries to "un-dim" the lights and "un-add" the noise to get the clear photo back.

2. The Problem with Old Models

Most previous AI models treated "making it bright" and "removing noise" as two different jobs. They would try to brighten the image first, then try to clean it. This is like trying to fix a leaky roof while the house is still on fire; the two problems interfere with each other.

3. The SADM Solution: The "Dimmer Switch"

The genius of this new model is a special ingredient called the Signal Attenuation Coefficient.

Imagine the AI has a magic dimmer switch built right into the "forward movie."

As the AI turns the lights down to create the "noise" version of the photo, it doesn't just turn the lights down; it simulates exactly how a real camera struggles in the dark. It makes the signal (the image) fade away faster than the noise appears.
Why does this matter? Because the AI learns that "Darkness" and "Noise" are best friends in this specific scenario. By teaching the AI that the image naturally fades into darkness while getting noisy, the AI learns to fix both problems at the exact same time when it plays the movie in reverse.

The Analogy:

Old Way: You have a muddy shoe. You wash it with water (makes it wet and muddy), then you scrub it (makes it clean but maybe damages the leather).
SADM Way: You realize the mud is the problem. You use a special cleaning solution that dissolves the mud and dries the leather simultaneously. You don't need two different tools; you just need one smart process that understands the mud and the leather are connected.

4. The "Pyramid" Trick (Speeding it Up)

Diffusion models are usually slow because they have to take thousands of tiny steps to clean the image (like peeling an onion one layer at a time).

The Fix: The authors use a Multi-Scale Pyramid. Imagine cleaning a huge room. Instead of cleaning every single inch of the floor one by one, you first clean the whole room quickly to get the big mess out (low resolution), and then you zoom in to clean the corners and details (high resolution).
This allows the AI to fix the photo in just 10 steps instead of 1,000, making it much faster without losing quality.

Why is this a Big Deal?

One Step, Two Wins: It fixes brightness and noise at the same time, so the photo looks natural, not blurry or weirdly colored.
No Extra Tools: Old methods needed extra "helper" networks to fix mistakes. This model does it all by itself, making it simpler and more efficient.
Real-World Results: The paper shows that this method produces clearer, brighter, and more detailed photos than any other current method, especially in tricky situations like night surveillance or driving in the dark.

In a nutshell: The authors built an AI that understands that "darkness" and "noise" go hand-in-hand. Instead of fighting them separately, it uses a special mathematical trick to fix both at once, resulting in crystal-clear photos from the darkest nights.

1. Problem Statement

Low-Light Image Enhancement (LLIE) faces a dual challenge: recovering brightness while suppressing noise and preserving fine details. Existing approaches suffer from significant limitations:

Traditional Methods: Rely on hand-crafted priors (e.g., Retinex theory) or histogram equalization, often leading to color distortion, noise amplification, and poor generalization in complex scenarios.
Deep Learning Methods: While CNNs and Transformers have improved performance, many adopt a multi-stage pipeline (e.g., "enhance-then-denoise" or "decompose-then-enhance"). This decoupling creates inconsistent optimization objectives, leading to suboptimal results, residual artifacts, and increased model complexity.
Diffusion Models: Although Diffusion Probabilistic Models (DDPMs) excel at modeling complex distributions, current diffusion-based LLIE methods largely retain the two-stage paradigm or rely on auxiliary correction networks. This fails to capture the intrinsic interdependence between illumination enhancement and noise suppression, compromising training and inference efficiency.

2. Methodology: Signal Attenuation Diffusion Model (SADM)

The authors propose SADM, a novel single-stage conditional diffusion framework that unifies brightness enhancement and noise suppression.

Core Innovation: Signal Attenuation Mechanism

Unlike standard DDPMs where the forward process adds noise to an image until it becomes pure Gaussian noise ( $x_T \sim \mathcal{N}(0, I)$ ), SADM introduces a signal attenuation coefficient ( $k_t$ ) into the forward process.

Forward Process: The transition is defined as $x_t = k_t(a_t x_{t-1} + b_t \epsilon_t)$ $x_{t} = k_{t} (a_{t} x_{t - 1} + b_{t} ϵ_{t})$ .
- As the diffusion steps ( $t$ ) increase, $k_t$ gradually attenuates the signal magnitude.
- By step $T$ , the image converges not just to noise, but to a state with zero signal intensity, physically simulating the degradation of a normal-light image into an extreme low-light condition.
Unified Objective: This mechanism forces the reverse denoising process to simultaneously learn brightness recovery (restoring signal intensity) and noise suppression (removing $\epsilon_t$ ) within a single network, eliminating the need for separate enhancement and denoising stages.

Mathematical Formulation

Variance Scheduling: The authors derive a recurrence constraint $k_{t-1}^4 a_t^2 + b_t^2 = k_t^2$ to ensure the second-moment (signal energy) of the image decays monotonically according to $E[\|x_t\|^2] = k_t^4$ . This aligns the diffusion process with the physical reality of low-light degradation.
Reverse Process: The reverse denoising network ( $f_\theta$ ) predicts the noise $\tilde{\epsilon}_t$ . The optimization objective minimizes the KL divergence between the true posterior and the learned distribution, transformed into a noise prediction loss weighted by the attenuation factors.

Architecture and Sampling

Network: Utilizes an SR3UNet as the backbone.
Conditions: The model takes the noisy image, time step, and adaptive image priors (specifically a dehazed version of the low-light image to reveal dark-region features and reduce color deviation) as joint inputs.
Sampling Efficiency: To address the slow inference of diffusion models, SADM employs:
- Multi-scale Pyramid Sampling: Processes images at decreasing resolutions to reduce computational load.
- DDIM Sampling: Uses Denoising Diffusion Implicit Models to accelerate the reverse process (reducing steps from 1000 to 10) without sacrificing quality.
Loss Function: A composite loss including:
- Noise Prediction Loss: L1 loss between predicted and ground-truth noise.
- Brightness Alignment Loss: A dynamic weighting of standard L1 loss and a brightness-aligned L1 loss (using a scaling factor $\lambda_{GT}$ ) to prevent color shifts and ensure luminance consistency with the Ground Truth.
- Perceptual Loss: VGG-19 feature matching to preserve visual coherence.

3. Key Contributions

Single-Stage Framework: SADM is the first diffusion model to unify brightness enhancement and noise suppression into a single-stage process by integrating a signal attenuation mechanism, removing the need for cascaded pipelines or correction modules.
Physical Priors: The introduction of the attenuation coefficient $k_t$ encodes the physical priors of low-light degradation directly into the diffusion formulation, providing a solid mathematical foundation for joint optimization.
Adaptive Prior Guidance: The use of a lightweight dehazing preprocessing step as a conditional input helps mitigate color deviation and improves the estimation of illumination in dark regions.
Efficiency: The combination of multi-scale pyramid sampling and DDIM allows for high-fidelity generation with significantly reduced inference time (10 steps).

4. Experimental Results

The method was evaluated on standard benchmarks: LOLv1, LOLv2-Real, and LOLv2-Syn, as well as unpaired datasets (DICM, NPE, MEF).

Quantitative Performance:
- LOLv1: Achieved state-of-the-art (SOTA) results with PSNR 28.0985 and SSIM 0.8770, outperforming all 12 compared methods (including Retinexformer, PyDiff, and Diff-Retinex++).
- LOLv2-Real: Ranked first in PSNR (28.9338) and second in SSIM.
- LOLv2-Syn: Delivered competitive PSNR (26.7042) and SSIM (0.9224).
Qualitative Performance:
- Visual comparisons show SADM effectively recovers details in dark regions while suppressing noise and avoiding the "smearing" or color distortion seen in competitors.
- It demonstrates superior generalization on unpaired real-world scenes (e.g., candlelight, outdoor night), maintaining natural tones without over-exposure.
Ablation Studies:
- Removing the attenuation coefficient ( $k_t$ ) caused a drastic drop in performance (PSNR dropped from 28.09 to 20.91), confirming the necessity of the signal attenuation mechanism.
- The combination of GT-mean loss and Perception loss yielded the best balance between pixel accuracy and visual quality.

5. Significance

This paper represents a paradigm shift in Low-Light Image Enhancement by moving away from heuristic, multi-stage pipelines toward a unified, physics-informed generative model.

Theoretical Impact: It demonstrates that diffusion models can be mathematically adapted to model specific physical degradation processes (signal attenuation) rather than just generic noise, offering a more interpretable and efficient approach to image restoration.
Practical Impact: By achieving SOTA performance with a single-stage architecture and efficient sampling, SADM offers a robust solution for critical applications like autonomous driving, surveillance, and remote sensing, where reliable low-light processing is essential.
Future Direction: The work suggests that integrating physical degradation models directly into the diffusion forward process is a promising avenue for solving other inverse problems in computer vision.