Naka-GS: Naka-Guided Chroma Correction and Point-Prior Compaction for Low-Light 3D Gaussian Splatting

arXiv:2604.11142 · cs.CV · Submitted 2026-04-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Naka-GS: Naka-Guided Chroma Correction and Point-Prior Compaction for Low-Light 3D Gaussian Splatting".

Jane: The gist: NAKA-GS is a bionics-inspired framework for low-light 3D Gaussian Splatting that jointly improves photometric restoration and geometric initialization.

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So, looking at "Naka-GS: Naka-Guided Chroma Correction and Point-Prior Compaction for Low-Light three dee Gaussian Splatting," what does this actually mean for the way we do this work <ref:2604.11142#pg1,for Low-Light 3D Gaussian Splatting>?

Jane: The authors mention that this approach can directly leverage mature low light image enhancement techniques and integrate them with existing reconstruction backbones in a modular manner.

Meng: They also point out that they designed mask guided supervision to target failure modes after enhancement, which helps mitigate issues like floating outliers or unstable geometry in weak texture regions.

Lu: The paper notes that this method starts with a Naka-guided chroma correction network that combines physics prior low light enhancement, dual branch input modeling, frequency decoupled correction, and mask guided optimization to suppress bright region chromatic artifacts and edge structure errors.

Tom: So the core idea is taking the image quality up first and then cleaning up the points before we start training our three dee Gaussian Splatting model <ref:2604.11142#pg1>.

Jane: That's right. It’s about sequential refinement, where each step cleans up what the previous one left behind to get a better final result for low light reconstruction.

Conclusion: Tom: So we’ve seen how Naka-GS handles low light by first fixing the colors and then cleaning up the points for Gaussian Splatting <ref:2604.11142#pg1>.

Jane: It really boils down to this paper creating one unified system instead of doing separate fixes for the pictures and the three dee shape.

Lu: They show that combining a physics-prior enhancement with targeted color correction, followed by point cloud preprocessing, works better than using different methods for each problem separately.

Meng: For us shipping models, that sequential refinement process is key because it improves input quality and then refines the geometry priors before the Gaussian training even starts.

Tom: The title itself points to how they’ve combined Naka-Guided Chroma Correction and Point-Prior Compaction to solve a specific problem in three dee Gaussian Splatting <ref:2604.11142#pg1>.

Jane: In simple terms, this paper suggests that if you’re dealing with low light, you don't just try to fix the pixels or just try to get the geometry right; you need a system that corrects both things sequentially and efficiently.

China University of Mining and Technology-Beijing

cs.CV

Submitted: 2026-04-13

Updated: 2026-10-14

Code: https://github.com/RunyuZhu/Naka-GS

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 91/100

The gist: The gist: NAKA-GS is a bionics-inspired framework for low-light 3D Gaussian Splatting that jointly improves photometric restoration and geometric initialization.

Key concepts

Naka–Rushton response algorithm
This physics-prior enhancement technique is used first to improve visibility in dark areas of an image. It models how light behaves under low-light conditions, providing a physically informed starting point for the 3D reconstruction process, which helps overcome severe visibility issues before learning color corrections.
Naka-Guided Chroma Correction
This is a neural network designed to fix color and brightness errors after initial enhancement. It uses dual inputs and frequency-decoupled correction to learn bounded corrections. This allows the model to suppress residual color artifacts while keeping important fine structures intact, leading to cleaner final images.
Point-Prior Compaction Module (PPM)
The PPM cleans and optimizes the dense point cloud generated during reconstruction. It aligns coordinates, aggregates nearby points via voxel pooling, and prunes redundant points adaptively. This process reduces the number of points from a million to about 60,000 without losing essential geometric information.
Frequency-decoupled correction
This technique in the chroma correction network separates image details into low and high frequencies. The model only applies color corrections to the low-frequency component, ensuring that sharp edges and fine textures—which are high-frequency details—are preserved during the refinement process.

Terminology

Summary

The gist: NAKA-GS is a bionics-inspired framework for low-light 3D Gaussian Splatting that jointly improves photometric restoration and geometric initialization.

How it works

  1. The overall pipeline consists of three stages: NAKA-based low-light enhancement, feed-forward multiview reconstruction, and Gaussian Splatting with pointcloud preprocessing <ref:2604.11142#pg8>.

  2. The first stage involves a physics-prior enhancement based on the Naka–Rushton response algorithm to improve visibility in dark regions <ref:2604.11142#pg4>.

  3. This is followed by a learnable Naka-guided chroma-correction network to refine the enhanced results <ref:2604.11142#pg6>.

Naka-Guided Chroma Correction

The Naka-guided chroma-correction model is designed to suppress specific photometric errors after initial enhancement <ref:2604.11142#pg6>. This network incorporates several advanced techniques:

** Dual-Branch Input construction:**

The input representation is constructed from a raw branch [I low, I naka, ∆] (4) and a normalized branch [eI low,eI naka, ∆e] (5) <ref:2604.11142#pg6>. This formulation allows the network to learn color and brightness correction more explicitly <ref:2604.11142#pg6>.

** ChromaGuided backbone and Frequency-decoupled correction:**

The network uses a U-Net-style encoder-decoder with three downsampling stages, a residual bottleneck with SE attention, and three upsampling stages with skip-connected fusion <ref:2604.11142#pg6>. Instead of directly regressing the output RGB image, it predicts a single-channel multiplicative correction map Mmul and a three-channel additive correction map Madd <ref:2604.11142#pg6>. To preserve high-frequency structures, the method employs frequency-decoupled correction by decomposing the Naka-enhanced image into low-frequency and high-frequency components, applying correction only to the low-frequency component while directly preserving the high-frequency residual <ref:2604.11142#pg6>.

Point Preprocessing Module (PPM)

The PPM addresses issues arising from noisy dense point clouds generated by feed-forward reconstruction <ref:2604.11142#pg4>. It follows a clean-before optimize strategy, refining the external dense point cloud in a minimally invasive preprocessing stage <ref:2604.11142#pg4>. The PPM performs three main tasks:

** Coordinate alignment:**

It first estimates a global transformation from camera centers and map the point cloud into the training coordinate space <ref:2604.11142#pg8>.

** Voxel pooling:**

This step is used to aggregate nearby samples into a more compact candidate point set <ref:2604.11142#pg8>.

** Distance-adaptive progressive pruning:**

This iterative process removes noisy and redundant points using a keep probability defined by P(ak) = min 1, dmin(ak) τ(t) + ϵ <ref:2604.11142#pg8>. The threshold is updated iteratively as τ(t+1) = τ(t) · exp β Mt / M0 <ref:2604.11142#pg8>.

Training Objective and Results

The training objective for the correction network is a compound loss: L = Lbase + Lgray + 0.8Lbright <ref:2604.11142#pg6>. The base loss includes terms like λrgbLrgb + λchromaLchroma + λssimLssim + λedgeLedge <ref:2604.11142#pg6>. Two mask-guided losses are introduced to target specific failure modes:

** Gray-edge mask loss:**

This term strengthens supervision on edges and texture-rich regions using a loss derived from the Laplacian response of the ground-truth grayscale image <ref:2604.11142#pg6>.

** Bright-region mask loss:**

This focuses on relatively bright regions in the prediction using a mask defined by Mbright(p) = Îgray(p) ≥ τ <ref:2604.11142#pg6>.

Experimentally, NAKA-GS achieved superior performance across nine low-light scenes, showing gains of +5.25 dB in PSNR over LITA-GS and consistently producing the best result on all nine scenes across all three metrics <ref:2604.11142#pg9>. The improvements are attributed to the sequential refinement process where photometric correction improves input quality, and PPM refines geometric priors before Gaussian initialization <ref:2604.11142#pg9>.

This unified framework improves restoration quality, initialization reliability, and training efficiency under challenging low-light conditions.

Improvements for AI systems

  1. The proposed Naka-guided chroma-correction model reduces bright-region color distortion and edge-structure errors without increasing inference complexity. This allows AI systems to generate 3D reconstructions from low-light images that maintain high photometric fidelity, specifically by suppressing bright-region chromatic artifacts and edge-structure errors.

  2. The Point Preprocessing Module (PPM) refines dense geometric priors through coordinate alignment, voxel pooling, and distance-adaptive progressive pruning. This enables the AI system to initialize 3D Gaussian Splatting with a cleaner and more compact geometric prior, leading to improved optimization stability and training efficiency by removing noisy outliers, locally over-dense clusters, and unstable geometry in weak-texture regions.

  3. The corrected images are then fed into a feed-forward multi-view reconstruction model to produce dense scene priors. This capability allows the system to generate high-quality dense scene priors, which serves as a more reliable input for subsequent Gaussian optimization, thereby improving the overall quality of novel view synthesis under dark conditions.

Sources

Related papers