Gaussian Light Field Splatting: A Physical Prior-Driven Vision Transformer for Unsupervised Low-Light Image Enhancement

2026-06-16 · Source: Computer Vision and Pattern Recognition · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Computer Vision · Depth: Expert, quick

Summary

Gaussian Light Field Splatting (GLFS) is a novel Vision Transformer designed for unsupervised low-light image enhancement, specifically addressing issues like local exposure imbalance and color distortion under complex, non-uniform illumination. Existing methods often struggle with these challenges, and most Vision Transformers lack explicit mechanisms for modeling physical illumination degradation priors. GLFS integrates continuous physical illumination modeling from Gaussian splatting into its Transformer architecture, representing scene illumination as a superposition of anisotropic Gaussian basis functions. It introduces physics-guided biases into self-attention to adaptively infer a spatial gain field, ensuring accurate and uniform restoration. To further enhance quality, GLFS employs a color-vector angular loss for hue consistency and a luminance-edge loss to improve structural fidelity of local details. Extensive evaluations confirm GLFS provides clear advantages in illumination correction and detail preservation, demonstrating leading performance.

Key takeaway

For Computer Vision Engineers developing unsupervised low-light image enhancement solutions, GLFS offers a new paradigm by integrating physical illumination priors into Vision Transformers. You should consider adopting its approach of modeling illumination with Gaussian basis functions and incorporating physics-guided biases into your self-attention mechanisms. This can significantly improve restoration uniformity and detail preservation, particularly under complex lighting conditions, by leveraging its novel color-vector angular and luminance-edge losses.

Key insights

GLFS integrates Gaussian splatting's physical illumination modeling into a Vision Transformer for unsupervised low-light image enhancement.

Principles

• Illumination can be modeled as anisotropic Gaussian basis functions.
• Physics-guided biases improve self-attention for spatial gain inference.
• Angular and edge losses enhance color and structural fidelity.

Method

GLFS represents scene illumination via anisotropic Gaussian basis functions, uses physics-guided biases in self-attention for spatial gain inference, and applies color-vector angular and luminance-edge losses.

In practice

• Apply Gaussian splatting principles to Vision Transformers.
• Use angular loss for hue consistency in image enhancement.
• Implement edge loss for structural detail preservation.

Topics

• Low-Light Image Enhancement
• Vision Transformers
• Gaussian Splatting
• Unsupervised Learning
• Illumination Modeling
• Computer Vision

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, Computer Vision Engineer

Related on AIssential

• AIGS-Net: Compact Illumination Field Modeling via 2D Gaussian Splatting for Fast Low-Light Image Enhancement — AIGS-Net uses adaptive 2D Gaussian Splatting and zero-parameter encoding for ultra-lightweight, fast low-light image enhancement.
• Fi-Gaussian: Frequency-Aware Implicit Gaussian Splatting for Single Image Dehazing — Fi-Gaussian uses frequency-aware implicit Gaussian splatting and physics-driven renormalization for state-of-the-art single image dehazing.
• EnerGS: Energy-Based Gaussian Splatting with Partial Geometric Priors — EnerGS uses a soft energy field from partial geometry to guide 3D Gaussian Splatting, improving reconstruction quality.
• Less Gaussians, Texture More: 4K Feed-Forward Textured Splatting — LGTM enables 4K novel view synthesis by decoupling geometry from resolution using textured Gaussian primitives.
• Voronoi-guided Bilateral 2D Gaussian Splatting for Arbitrary-Scale Hyperspectral Image Super-Resolution — GaussianHSI uses Voronoi-guided 2D Gaussian splatting for arbitrary-scale hyperspectral image super-resolution.
• WildSplatter: Feed-forward 3D Gaussian Splatting with Appearance Control from Unconstrained Images — WildSplatter enables fast, pose-free 3DGS from unconstrained images by disentangling geometry and appearance via learned embeddings.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.