Feature Request | Low-Light Image Enhancement

References for existing research papers:

1. Big picture


2. Low‑light image enhancement (general)

2.1 Retinex‑based Transformers

Retinexformer: One-stage Retinex-based Transformer for Low-light Image Enhancement – ICCV 2023

  • Key idea: Formulates a One‑stage Retinex‑based Framework (ORF): first estimates illumination to “light‑up” the image, then restores corruptions (noise, artifacts) with an Illumination‑Guided Transformer (IGT) that uses illumination representations to guide non‑local interactions across regions with different lighting.
  • Status: First Transformer‑based method for LLIE; significantly outperforms prior SOTA on 13 benchmarks and shows gains in low‑light object detection. Code/models widely used as baseline; second place in NTIRE 2024 Low Light Enhancement Challenge.

LightingFormer: Transformer-CNN hybrid network for low-light image enhancement – Computers & Graphics 2024

  • Key idea: Proposes a Transformer–CNN hybrid block with mixed attention: Transformer models long‑range dependencies, CNN extracts low‑level local features; uses a U‑Net discriminator for adaptive region‑wise enhancement and to suppress over/underexposure and noise.
  • Status: Reported to outperform prior SOTA quantitatively and qualitatively, with improved downstream object detection.

2.2 SNR‑aware models

SNR-Aware Low-Light Image Enhancement – CVPR 2022

  • Key idea: Introduces an SNR‑aware Transformer that uses a signal‑to‑noise ratio prior map to guide feature fusion and avoid tokens from very low‑SNR regions; combines with CNN for spatially varying enhancement.
  • Status: Consistently better than prior SOTA on 7 benchmarks; validated by large‑scale user study.

Unsupervised Low Light Image Enhancement Using SNR-Aware Swin Transformer – 2023

  • Key idea: Dual‑branch Swin Transformer network guided by an SNR prior map; uses unsupervised Retinex‑based losses so training does not require paired data.
  • Status: Competitive with baselines while reducing reliance on paired supervision.

SMT: SNR-Aware Mamba-Transformer for Low-Light Image Enhancement – PRCV 2025

  • Key idea: Introduces an SNR‑aware Mamba module for global noise modeling, an eight‑directional Mamba scanning mechanism for multi‑perspective context, and a multi‑scale attention aggregation module for fine detail recovery.
  • Status: Outperforms or matches SOTA on 8 benchmarks with better visual quality.

2.3 Efficient / self‑calibrated illumination learning

Toward Fast, Flexible, and Robust Low-Light Image Enhancement (SCI / SCI++) – CVPR 2022

  • Key idea: Self‑Calibrated Illumination learning: cascaded illumination stages with weight sharing, plus a self‑calibrator so only a single basic block is used at inference. SCI++ further disentangles self‑calibrator and illumination estimator for better convergence and interpretability.
  • Status: Very strong trade‑off between quality and speed; unsupervised loss improves generalization to in‑the‑wild scenes.

LLFormer: Ultra-High-Definition Low-Light Image Enhancement: A Benchmark and Transformer-Based Method – AAAI 2023

  • Key idea: Proposes a 4K/8K UHD‑LOL benchmark and LLFormer, whose core components are axis‑based multi‑head self‑attention (height/width axes sequentially) and cross‑layer attention fusion to keep complexity linear.
  • Status: Outperforms prior SOTA on UHD‑LOL and standard LOL datasets; widely used as UHD LLIE baseline.

2.4 Color space / decoupling strategies

HVI: A New Color Space for Low-Light Image Enhancement / HVI-CIDNet+ – 2025

  • Key idea: Proposes Horizontal/Vertical‑Intensity (HVI) color space designed to reduce red/black artifacts common in RGB and HSV; introduces Color and Intensity Decoupling Network (CIDNet) with a Lighten Cross‑Attention (LCA) block to learn accurate photometric mapping in HVI space.
  • Status: Outperforms SOTA on 10 datasets; HVI‑CIDNet is core of 1st place solution in NTIRE 2025 Low Light Image Enhancement Challenge.

Low-light Image Enhancement Algorithm Based on Transformer – SPIC 2024

  • Key idea: Redesigns LLIE network using Transformer: extracts channel‑wise self‑attention features, fuses with shallow features, then refines; emphasizes efficiency and global dependencies vs traditional Retinex/CNN pipelines.
  • Status: Representative 2024 Transformer LLIE work with improved performance and efficiency over prior CNN‑based methods.

CF-UFormer: Low-light image enhancement using transformer with color fusion – 2024

  • Key idea: U‑shape Transformer with color fusion module to better handle color and luminance jointly; improves visual quality and stability under severe low‑light conditions.

2.5 Surveys & benchmarks (LLIE)

  • A Survey of Deep Learning-Based Low-Light Image Enhancement – Sensors 2023
    Systematic classification of DL‑based LLIE methods, datasets, and metrics; summarizes representative CNN/GAN/Transformer models and their trade‑offs.
  • A survey on image enhancement for Low-light images – 2023
    Focuses on DL‑based vs traditional methods and data characteristics.
  • NTIRE 2024 Challenge on Low Light Image Enhancement: Methods and Results – CVPRW 2024
    Challenge overview; top solutions set practical SOTA on standard LOL and real‑world datasets.
  • NTIRE 2026 Challenge on Efficient Low Light Image Enhancement (E‑LLIE) – 2026
    Overview of efficient LLIE methods; emphasizes lightweight, mobile‑friendly architectures.

3. Shadow removal

3.1 Global context Transformers

ShadowFormer: Global Context Helps Shadow Removal – AAAI 2023

  • Key idea: Derives a Retinex‑based shadow model and proposes Shadow‑Interaction Module (SIM) with Shadow‑Interaction Attention (SIA) to model context correlation between shadow and non‑shadow regions in a multi‑scale channel attention Transformer.
  • Status: Achieves SOTA on ISTD, ISTD+, SRD with up to 150× fewer parameters than prior methods.

CRFormer: A cross-region transformer for shadow removal – 2024

  • Key idea: Introduces region‑aware cross‑attention to transfer features from non‑shadow to shadow regions, improving boundary consistency and reducing artifacts.
  • Status: Strong performer on standard shadow datasets; cross‑region design is influential.

3.2 Mask‑free / FFT‑based methods

ShadowRefiner: Towards Mask-free Shadow Removal via Fast Fourier Transformer – CVPRW 2024

  • Key idea: Mask‑free shadow removal and refinement via Fast Fourier Transformer; learns mappings in both spatial and frequency domains using Fast‑Fourier Attention (FFAT) to reduce pixel misalignment and refine details.
  • Status: 1st place in NTIRE 2024 Shadow Removal Perceptual Track, 2nd in Fidelity Track.

3.3 Diffusion & hybrid models

ShadowFormer++: multi-scale shadow priors and diffusion refinement – 2025

  • Key idea: Combines Transformer + diffusion: Multi‑Scale Local Shadow Perception (MS‑LSPM), Shadow‑Aware Transformer Encoder (SATE), and Diffusion‑Inspired Refinement Module (DIRM) for progressive shadow‑free reconstruction.
  • Status: Achieves SOTA MAE/PSNR/SSIM on ISTD/ISTD+/SRD.

3.4 Surveys & benchmarks (shadows)

  • Single-Image Shadow Removal Using Deep Learning: A Comprehensive Survey – 2024
    Detailed taxonomy of shadow removal networks, losses, and datasets; summarizes quantitative/qualitative comparisons.
  • Unveiling Deep Shadows: A Survey and Benchmark on Image and Video Shadow Detection, Removal, and Generation – IJCV 2026
    Unified survey and benchmark for shadow detection/removal/generation; re‑trains representative methods under standardized settings and highlights cross‑dataset generalization issues.
  • Awesome-Shadow-Removal – curated list of supervised/unsupervised deep shadow methods.

4. Specular / highlight removal

4.1 Attention‑based highlight removal

Dual-Hybrid Attention Network for Specular Highlight Removal (DHAN-SHR) – ACM Multimedia 2024

  • Key idea: End‑to‑end network with Adaptive Local Hybrid‑Domain Dual Attention Transformer (L‑HD‑DAT) and Adaptive Global Dual Attention Transformer (G‑DAT); models inter‑channel/pixel and spectral‑domain dependencies without extra priors.
  • Status: Outperforms 18 SOTA methods on a large‑scale benchmark, setting a new standard for specular highlight removal.

4.2 Coarse‑to‑fine + diffusion refinement

Document Specular Highlight Removal with Coarse-to-Fine Strategy – ICDAR 2024

  • Key idea: Three‑module framework:
    • Coarse Predictor (CP): Transformer‑based U‑Net recovers main content.
    • Global Discriminator (GD): Ensures global consistency of coarse result.
    • Refinement Predictor (RP): Conditional Denoising Diffusion Probabilistic Model (DDPM) predicts residual between GT and CP output.
  • Status: Surpasses SOTA on four highlight benchmarks.

4.3 Privacy‑preserving / federated learning

Specular highlight removal by federated generative adversarial network with attention mechanism (FL-AttGAN) – Scientific Reports 2024

  • Key idea: Uses federated learning to train a global highlight‑removal model without centralizing sensitive images; combines AttGAN with attention mechanisms to focus on highlight regions and improve realism.
  • Status: Outperforms existing methods on SD1/SD2/RD datasets under privacy‑preserving constraints.

Hi.

Interesting, would be nice if possible to see a couple of examples also, how do think this would compare with a combination of using Denoise MAX and one of the Adjust Lighting Models Denoise MAX for reducing extreme noise and Adjust Lighting to brighten the shadows yet retaining the highlights

The main concept here is to recover highlight and shadow clipping, where data is lost entirely, and therefore would have to be estimated by a generative model. Any other model that attempts to recover existing details would fail at this task.

I am listing prior work here for reference, hopefully Topaz could use its magic with NeuroServer and deliver the goods but at much, much faster processing times.