Research map / Image generation / Image restoration
Empowering Resampling Operation for Ultra-High-Definition Image Enhancement with Model-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 LMAR, which learns model-aware compensatory kernels and steganographically embeds them into downscaled UHD inputs so that interpolation resampling becomes consistent with a frozen enhancer's features, improving UHD low-light enhancement at low resolutions.
Problem
Downsampling-enhancement-upsampling pipelines for UHD image enhancement disentangle interpolation resamplers from the enhancer, losing information the model relies on and yielding sub-optimal results.
Approach
Dynamically predicts image-specific compensatory convolution kernels (via implicit neural representation with scale features and back-projection) and embeds the derived compensatory information into the downscaled input, trained with feature-consistency, attention, and adversarial losses against a frozen pre-trained enhancer, while remaining compatible with any interpolation resampler.
Key results
- Up to 1.2 dB PSNR gain at x1.8 resampling scale on UHD-LOL4K (abstract).
- On UHD-LOL4K with CNN-I, bicubic at x4 (scale 540x960 region) improves from about 25.40 to 26.36 PSNR under the full LMAR setup in the ablation (Table 5, x4 scale: baseline 25.40, full 26.36).
- Gains reported across five classical resizers and multiple backbones (CNN-I, CNN-H, Restormer) with only a minor parameter increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply LMAR to Bilinear resampling | CNN-I · UHD-LOL4K, 540×960 (×4) | PSNR: 25.90 -> 26.27 | Helped |
| Apply LMAR to Bilinear resampling | CNN-H · 4KIL, 540×960 (×4) | PSNR: 26.38 -> 26.57 | Helped |
| Apply LMAR to Bilinear resampling | Restormer · 4KIL, 540×960 (×4) | PSNR: 26.71 -> 26.77 | Helped |
| Use bottleneck features instead of front features as the feature constraint (ablation) | LMAR on CNN-I · UHD-LOL4K, ×2 | PSNR: 29.98 -> 30.20 | Helped |
| Use bottleneck features instead of front features as the feature constraint (ablation) | LMAR on CNN-I · UHD-LOL4K, ×4 | PSNR: 25.86 -> 26.36 | Helped |
| Use bottleneck features instead of end features as the feature constraint (ablation) | LMAR on CNN-I · UHD-LOL4K, ×2 | PSNR: 30.05 -> 30.20 | Helped |
| Use bottleneck features instead of end features as the feature constraint (ablation) | LMAR on CNN-I · UHD-LOL4K, ×4 | PSNR: 25.50 -> 26.36 | Helped |
| Use the discriminator instead of VGG perceptual loss for distribution restriction (ablation) | LMAR on CNN-I · UHD-LOL4K, ×2 | PSNR: 30.03 -> 30.20 | Helped |
| Use the discriminator instead of VGG perceptual loss for distribution restriction (ablation) | LMAR on CNN-I · UHD-LOL4K, ×4 | PSNR: 25.47 -> 26.36 | Helped |
Surprising findings
- Manipulating the interpolation kernel alone can make results either better or worse, showing high model sensitivity to the resampler.
Limitations
- Requires training an enhancer per dataset and a collaborative training stage per enhancer
- Evaluated only on low-light UHD enhancement datasets
Experimental setup
Datasets: UHD-LOL4K 4KIL
Benchmarks: UHD-LOL4K 4KIL
Metrics: PSNR SSIM
Baselines: Nearest Bilinear Bicubic Lanczos2 Lanczos3 Task-aware image downsampling (TAD)
Builds on: LIIF Task-aware image downsampling (TAD) back-projection Restormer
Closest papers by meaning
- AutoLUT: LUT-Based Image Super-Resolution with Automatic Sampling and Adaptive Residual Learning (CVPR 2025 · similarity 0.83)
- FlowIE: Efficient Image Enhancement via Rectified Flow (CVPR 2024 · similarity 0.82)
- UltraPixel: Advancing Ultra High-Resolution Image Synthesis to New Peaks (NeurIPS 2024 · similarity 0.82)
- ResMaster: Mastering High-Resolution Image Generation via Structural and Fine-Grained Guidance (AAAI 2025 · similarity 0.81)
- Low-Light Image Enhancement via Generative Perceptual Priors (AAAI 2025 · similarity 0.81)
- AGLLDiff: Guiding Diffusion Models Towards Unsupervised Training-free Real-world Low-light Image Enhancement (AAAI 2025 · similarity 0.80)
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.