Research map / Vision-language and multimodal / Human and embodied modeling
Learning Adaptive Lighting via Channel-Aware Guidance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes LALNet, a multi-task channel-aware network that fuses color-separated and color-mixed features via Light Guided Attention to handle exposure correction, image retouching, low-light enhancement, and tone mapping.
Problem
Light-related tasks (exposure correction, image retouching, low-light enhancement, tone mapping) are usually solved individually, and existing unified frameworks underperform task-specific methods because they ignore channel-specific light properties.
Approach
Exploits per-channel light differences in spatial and frequency domains: a dual-domain channel modulation branch extracts color-separated features, a mixed-channel branch with wavelet modulation and a light state space module extracts color-mixed features, and Light Guided Attention uses the separated features as queries to guide the mixed features, with iterative detail enhancement over a learnab
Key results
- On SCIE exposure correction, LALNet reaches 23.80 average PSNR and 0.7372 SSIM, improving over CoTF by 2.29 dB PSNR in the paper's claim.
- On HDR+ image retouching, LALNet achieves 30.24 PSNR versus 28.09 for MambaIR, with LALNet-Tiny at 230K parameters still reaching 29.30.
- On LOLv1 low-light enhancement, LALNet obtains 25.26 PSNR and 0.855 SSIM at 6.70 GFLOPs, outperforming RetinexMamba (24.03 PSNR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MCM instead of a convolution block (ablation) | LALNet · SCIE exposure correction | PSNR: 23.95 -> 24.63 | Helped |
| add DDCM instead of group convolution for channel-separated features (ablation) | LALNet · SCIE exposure correction | PSNR: 23.47 -> 24.63 | Helped |
| add LGA rather than directly summing channel-mixed and channel-separated features (ablation) | LALNet · SCIE exposure correction | PSNR: 23.53 -> 24.63 | Helped |
| add LSSM instead of residual blocks (ablation) | LALNet · SCIE exposure correction | PSNR: 22.81 -> 24.63 | Helped |
| use 3 rather than 2 pyramid levels (ablation) | LALNet · SCIE exposure correction | PSNR: 23.45 -> 24.63 | Helped |
| use 4 rather than 3 pyramid levels (ablation) | LALNet · SCIE exposure correction | PSNR: 24.63 -> 23.07 | Hurt |
| use 3 rather than 2 pyramid levels (ablation) | LALNet · SCIE exposure correction | TMQI: 0.8735 -> 0.8667 | Hurt |
| use 3 rather than 2 pyramid levels (ablation) | LALNet · SCIE exposure correction | LPIPS: 0.1240 -> 0.1270 | Hurt |
| use 3 rather than 2 pyramid levels (ablation) | LALNet · SCIE exposure correction | ΔE: 8.18 -> 7.53 | Helped |
| replace LSSM with Self-Attention (ablation) | LALNet · SCIE exposure correction | PSNR: 24.62 -> 24.41 | Hurt |
| replace LSSM with ResBlock (ablation) | LALNet · SCIE exposure correction | PSNR: 24.62 -> 22.81 | Hurt |
| use group convolution instead of traditional convolution (ablation) | LALNet · SCIE exposure correction | PSNR: 23.98 -> 24.62 | Helped |
Surprising findings
- Replacing the SS2D module with a Self-Attention block gives comparable results (24.41 vs 24.62 PSNR in the appendix ablation), suggesting the framework design rather than SS2D drives gains.
Limitations
- Paper mentions no formal limitations section; the LGA and DDCM gains rest on empirical ablation only.
- Exposure correction table shows LALNet with weaker LPIPS (0.1397) than LALNet-Tiny (0.1280), so gains are not uniform across metrics.
Experimental setup
Datasets: SCIE/SICE HDR+ Burst Photography LOL HDRI Haven MSEC HDR Survey UVTM
Benchmarks: SCIE HDR+ LOL-v1 HDRI Haven MSEC HDR Survey UVTM
Metrics: PSNR SSIM TMQI LPIPS deltaE NIQE MUSIQ GFLOPs
Baselines: CoTF Retinexformer RetinexMamba MambaIR Restormer CSRNet HDRNet DeepLPF CLUT SepLUT LLFLUT FECNet
Builds on: Retinexformer RetinexMamba CoTF MambaIR SS2D
Closest papers by meaning
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.85)
- Lux Post Facto: Learning Portrait Performance Relighting with Conditional Video Diffusion and a Hybrid Dataset (CVPR 2025 · similarity 0.85)
- LumiNet: Latent Intrinsics Meets Diffusion Models for Indoor Scene Relighting (CVPR 2025 · similarity 0.84)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.84)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.84)
- DreamLight: Towards Harmonious and Consistent Image Relighting (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.