💻 computer science

Light 'em Up: Enabling Few-Shot Low-Light 3D Gaussian Splatting with Multi-Scale Explicit Retinex Illumination Decoupling

The paper introduces MERID-GS, a multi-scale explicit Retinex illumination-decoupled framework that leverages 3D Gaussian Splatting to achieve state-of-the-art, few-shot low-light 360° novel view synthesis by effectively separating illumination and reflectance while suppressing noise and ensuring view consistency.

Original authors: YuHao Yin, Zongji Wang, Yuanben Zhang, Biqing Li, Jiesong Bai, Junyi Liu

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: YuHao Yin, Zongji Wang, Yuanben Zhang, Biqing Li, Jiesong Bai, Junyi Liu

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 take a beautiful, 360-degree panoramic photo of a garden at night. You have a few photos from different angles, but they are dark, grainy, and full of "snow" (noise). If you try to stitch these together to create a 3D model that you can walk around in, standard computer programs usually fail. They get confused by the darkness, turning the garden into a blurry, color-shifting mess, or they amplify the noise until the image looks like static on an old TV.

This paper introduces a new method called MERID-GS (Light 'em Up) that solves this problem. Here is how it works, using simple analogies:

1. The Problem: The "Dark Room" Confusion

Think of a standard 3D camera system like a painter trying to recreate a scene. If the painter is handed a photo that is too dark, they might guess that a shadow is actually a black wall, or that the grainy noise is a real texture. When they try to paint the scene from a new angle (one they haven't seen before), their guesses fall apart, and the colors look wrong.

2. The Solution: Separating the "Flash" from the "Paint"

The core idea of MERID-GS is based on an old theory called Retinex. Imagine looking at a red apple in a dark room.

Illumination: This is the "flash" or the light hitting the apple.
Reflectance: This is the "paint" or the actual red color of the apple itself.

Most computer programs mix these two together. MERID-GS acts like a smart editor that peels them apart.

It says, "Okay, this part is just the darkness (Illumination). Let's fix that."
It says, "And this part is the actual object (Reflectance). Let's keep that pure."

By separating the light from the object, the computer can brighten the dark areas without messing up the true colors of the objects.

3. The "Noise Filter" (IS-FGA)

When you brighten a dark photo, you often accidentally make the "snow" or graininess worse.

The Analogy: Imagine trying to hear a whisper in a windy room. If you turn up the volume, you hear the whisper, but the wind noise gets louder too.
The Fix: MERID-GS uses a special tool called Illumination-State-Guided Frequency Gated Attention. Think of this as a "smart noise-canceling headphone" for images. It listens to the "wind" (noise) and the "whisper" (the actual structure of the object) separately. It turns down the volume on the wind but keeps the whisper loud and clear, especially in the darkest corners of the image.

4. The "Quick-Change Artist" (Few-Shot Adaptation)

Usually, teaching a computer to understand a new scene requires showing it hundreds of photos and training it for days.

The Analogy: Imagine a chef who usually needs to taste a soup 100 times to get the recipe right. MERID-GS is like a master chef who can taste the soup just 10 times and immediately know exactly how to season it for a new batch.
How it works: The system learns a general "recipe" for how light works. When it sees a new, dark scene, it only needs a tiny handful of photos (about 10) to quickly adjust its "seasoning" (the Reflection Head) and get it right. It can do this entire process in about 5 minutes.

5. The Result: A 3D World from the Dark

Once the system cleans up the photos and separates the light from the objects, it uses a technique called 3D Gaussian Splatting.

The Analogy: Instead of building a solid 3D model out of bricks, imagine the scene is made of millions of tiny, glowing, fuzzy balls (Gaussians). The computer arranges these balls to form the shape of the garden.
Because the input photos were cleaned up and the colors were fixed, these "fuzzy balls" arrange themselves perfectly. The result is a high-quality, 360-degree view that looks real, even though the original photos were taken in the dark.

Summary of Claims

What it does: It creates 3D, 360-degree views of scenes taken in low light.
How fast: It takes about 5 minutes to adapt to a new scene.
How much data: It only needs about 10 photos to work.
The Dataset: The authors also created a new set of test data (called LLD) specifically for this task, covering full 360-degree scenes in the dark, because existing data wasn't good enough.
The Performance: In tests, this method produced clearer, more accurate, and more consistent 3D views than any other method currently available, especially when moving to new, unseen scenes.

In short, MERID-GS is a tool that lets computers "see" clearly in the dark, separate the light from the objects, and build a perfect 3D world from just a few blurry photos.

1. Problem Statement

Low-light 360° Novel View Synthesis (NVS) remains a significant challenge in computer vision. Existing methods struggle with three primary issues when applied to real-world, low-exposure scenes:

Photometric Inconsistency: Insufficient illumination leads to noise amplification and color distortion, disrupting the appearance consistency required for multi-view geometry reconstruction.
View-Dependent Artifacts: Unsupervised approaches often suffer from severe color drift when viewpoints vary significantly (full 360° coverage), while supervised 2D enhancement models fail to generalize to unseen 3D scenes without retraining.
Data Scarcity: Most existing low-light NVS methods require dense view inputs or extensive paired training data, making them impractical for sparse-view, few-shot scenarios common in real-world applications (e.g., nighttime autonomous driving, cultural heritage digitization).

2. Methodology: MERID-GS

The authors propose MERID-GS, a framework that integrates Multi-Scale Explicit Retinex Illumination Decoupling with 3D Gaussian Splatting (3DGS). The pipeline consists of three distinct stages:

A. Multi-Scale Explicit Retinex Illumination Decoupling (MERID)

Instead of modeling illumination degradation implicitly within the 3D representation, MERID-GS performs explicit decoupling in the image domain based on Retinex theory ( $I = L \times R$ , where $I$ is intensity, $L$ is illumination, and $R$ is reflectance).

Learnable Regularized Gain: To avoid numerical instability in the logarithmic domain (common in low light), the model operates in the spatial domain. It introduces a learnable gain function $g(x)$ that adapts to luminance and structural information, suppressing noise amplification while enhancing dark regions.
Illumination-State Representation: The model constructs an illumination-state feature ( $\mathbf{F}_{illu}$ ) by encoding the low-pass luminance, reflectance, and log-gain. This serves as a prior for subsequent denoising.

B. Illumination-State-Guided Frequency Gated Attention (IS-FGA)

To address the entanglement of noise and true structures in low-light features, the authors introduce a novel attention mechanism within a U-shaped encoder-decoder architecture:

Frequency-Conditioned Value Modulation: The value ( $V$ ) branch of the attention mechanism is decomposed into low, mid, and high-frequency bands. These bands are modulated based on the illumination-state feature, allowing the model to selectively enhance structural frequencies while suppressing noise-dominated frequencies.
Illumination-State-Conditioned Gating: An element-wise gating mechanism further suppresses noise propagation in dark regions before standard attention aggregation occurs.

C. Reflection Head & 3D Gaussian Splatting

Reflection Head: A lightweight module (using $1\times1$ convolutions and GELU) is applied to the reflectance component to correct cross-scene color and material deviations. Crucially, this head is designed for few-shot adaptation, requiring only a few samples from a new scene to fine-tune.
3DGS Synthesis: The corrected reflectance images are fed into 3D Gaussian Splatting. Because the illumination has been explicitly decoupled and normalized, the 3DGS optimizer can focus on geometry and reflectance, leading to stable convergence even with sparse inputs.

3. Key Contributions

MERID-GS Framework: The first method to combine explicit Retinex-based illumination decoupling with 3DGS for low-light NVS, enabling high-quality rendering on unseen scenes from sparse observations.
IS-FGA Mechanism: A novel attention module that uses illumination states to guide frequency gating, effectively suppressing noise propagation while preserving structural details in dark regions.
Few-Shot Adaptability: The system achieves stable synthesis with only 10 normal-light input views for adaptation, reducing per-scene training time to approximately 5 minutes.
LLD Dataset: The authors constructed LLD (Low-Light Dataset), a full 360° multi-view dataset covering 6 scenes with paired low-light and normal-light images, addressing the lack of benchmarks for this specific task.

4. Experimental Results

The method was evaluated on NeRF360, LOM_full, and the new LLD dataset, comparing against state-of-the-art unsupervised (e.g., Luminance-GS) and supervised (e.g., Retinexformer, MambaLLIE) baselines.

Quantitative Performance:
- MERID-GS achieved SOTA performance across all metrics (PSNR, SSIM, LPIPS).
- Average PSNR: 25.03 dB (vs. ~17–19 dB for the next best supervised methods).
- SSIM: 0.796 (vs. ~0.67 for baselines).
- LPIPS: 0.166 (lower is better, significantly outperforming baselines).
Qualitative Performance:
- Visual comparisons show MERID-GS eliminates the color drift and noise amplification seen in other methods.
- It preserves geometric consistency and fine details in dark regions where other methods fail or produce artifacts.
Efficiency:
- The few-shot adaptation process takes only ~5 minutes per scene.
- Ablation studies confirmed that removing the Reflection Head or IS-FGA significantly degrades performance (dropping PSNR by ~7 dB and ~0.55 dB, respectively).

5. Significance

This work bridges the gap between 2D low-light enhancement and 3D scene reconstruction. By explicitly decoupling illumination from reflectance, MERID-GS solves the fundamental issue of "illuminating" a 3D scene without confusing lighting changes with geometric or material changes.

Practical Impact: It enables high-fidelity 3D reconstruction in challenging environments (nighttime, indoor low-light) with minimal data requirements, making it viable for applications like autonomous driving, AR/VR, and digital heritage preservation.
Generalization: The explicit illumination modeling allows the model to generalize to completely unseen scenes without the need for extensive retraining, a critical step forward for real-world deployment.