Research map / Image generation / Image restoration
VQCNIR: Clearer Night Image Restoration with Vector-Quantized Codebook
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 VQCNIR, a night image restoration network that uses a pretrained VQGAN codebook as high-quality prior, with an adaptive illumination enhancement module (AIEM) and deformable bi-directional cross-attention (DBCA) to align degraded features with codebook features.
Problem
Restoring night images that suffer from joint low illumination and motion blur, where handcrafted or low-quality-derived priors are unreliable and codebook features mismatch degraded features in illumination and structure.
Approach
Use a VQGAN codebook trained on normal-light clean images as a data-driven prior, adding an illumination-curve module (AIEM, Zero-DCE-style curves estimated from inter-channel dependencies) to match illumination and a parallel decoder with bi-directional cross-attention plus deformable convolution (DBCA) to fuse codebook and degraded features.
Key results
- On LOL-Blur, VQCNIR reaches 27.79 dB PSNR, 0.875 SSIM and 0.096 LPIPS, outperforming the next-best LEDNet (25.74 dB) by 2.05 dB.
- On Real-LOL-Blur, VQCNIR achieves MUSIQ 51.04, NRQM 7.064 and NIQE 4.599, the best on all three non-reference metrics (LEDNet MUSIQ 39.11).
- Ablation shows VQGAN alone gives 10.79 dB PSNR, Decoder D setting 26.58 dB, adding AIEM 27.48 dB and full model with DBCA 27.79 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add decoder D to the VQGAN codebook prior (ablation) | VQCNIR · LOL-Blur | PSNR: 10.79 -> 26.58 | Helped |
| Add decoder D to the VQGAN codebook prior (ablation) | VQCNIR · LOL-Blur | SSIM: 0.3028 -> 0.8486 | Helped |
| Add DBCA to decoder D (ablation) | VQCNIR · LOL-Blur | PSNR: 26.58 -> 26.89 | Helped |
| Add DBCA to decoder D (ablation) | VQCNIR · LOL-Blur | SSIM: 0.8486 -> 0.8599 | Helped |
| Add AIEM to decoder D (ablation) | VQCNIR · LOL-Blur | PSNR: 26.58 -> 27.48 | Helped |
| Add AIEM to decoder D (ablation) | VQCNIR · LOL-Blur | SSIM: 0.8486 -> 0.8692 | Helped |
Surprising findings
- Directly reconstructing night images with a VQGAN trained on normal-light images yields only 10.79 dB PSNR, showing codebook priors do not transfer without alignment.
Limitations
- Real-world evaluation relies only on non-reference metrics since no ground truth exists for Real-LOL-Blur.
- Experiments are limited to the LOL-Blur dataset for training.
Experimental setup
Datasets: LOL-Blur
Benchmarks: LOL-Blur Real-LOL-Blur
Metrics: PSNR SSIM LPIPS MUSIQ NRQM NIQE
Baselines: LEDNet Restormer MIMO LLFlow Uformer DMPHN DeblurGAN-v2 KinD++ DRBN Zero-DCE
Builds on: VQGAN Zero-DCE NAFNet LEDNet
Closest papers by meaning
- NightHaze: Nighttime Image Dehazing via Self-Prior Learning (AAAI 2025 · similarity 0.87)
- DarkIR: Robust Low-Light Image Restoration (CVPR 2025 · similarity 0.86)
- Unveiling Details in the Dark: Simultaneous Brightening and Zooming for Low-Light Image Enhancement (AAAI 2024 · similarity 0.85)
- Luminance-Aware Statistical Quantization: Unsupervised Hierarchical Learning for Illumination Enhancement (NeurIPS 2025 · similarity 0.85)
- Zero-Reference Low-Light Enhancement via Physical Quadruple Priors (CVPR 2024 · similarity 0.85)
- Interpretable Unsupervised Joint Denoising and Enhancement for Real-World low-light Scenarios (ICLR 2025 · similarity 0.85)
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.