Research map / Image generation / Image restoration
From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective
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 ERR, a three-stage UHD image restoration network that decouples restoration into zero-frequency enhancement, low-frequency restoration, and high-frequency refinement guided by a DCT-based spectral analysis, using a frequency-windowed KAN for the final stage.
Problem
Ultra-high-definition image restoration is hard due to ultra-high resolution, complex content and fine details, and existing downsample-enhance-upsample methods lose critical information.
Approach
Analyze restoration via DCT spectrum swapping to show the zero-frequency (global) component matters first, then low-frequency (coarse) and high-frequency (fine) content; implement three coarse-to-fine sub-networks (transformer on 8x downsampled features, Mamba on 4x features, and frequency-windowed KAN at full resolution) with per-stage frequency-band losses.
Key results
- On UHD-LL low-light enhancement, ERR reaches 27.57 PSNR / 0.932 SSIM with 1.131M parameters, above UHDFormer (27.11) and Wave-Mamba (27.35).
- On 4K-Rain13k deraining, ERR achieves 34.48 PSNR vs. 34.28 for UDR-Mixer, with 0.952 SSIM.
- On UHD-Haze dehazing, ERR reports 25.12 PSNR, a 2.53 dB gain over UHDFormer (22.59).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ERR instead of Wave-Mamba | UHD-LL low-light enhancement · UHD-LL | PSNR: 27.35 -> 27.57 | Helped |
| Use ERR instead of UDR-Mixer | 4K-Rain13k image deraining · 4K-Rain13k | PSNR: 34.28 -> 34.48 | Helped |
| Use ERR instead of UHDFormer | UHD-Haze image dehazing · UHD-Haze | PSNR: 22.59 -> 25.12 | Helped |
| Use ERR instead of UHDFormer | UHD-Blur image deblurring · UHD-Blur | PSNR: 28.82 -> 29.72 | Helped |
| Set frequency cutoff k to 32 rather than 64 (ablation) | ERR · UHD-LL | PSNR: 27.57 -> 26.71 | Hurt |
| Set frequency cutoff k to 96 rather than 64 (ablation) | ERR · UHD-LL | PSNR: 27.57 -> 27.52 | Hurt |
| Set frequency cutoff k to 128 rather than 64 (ablation) | ERR · UHD-LL | PSNR: 27.57 -> 27.29 | Hurt |
| Set frequency cutoff k to 256 rather than 64 (ablation) | ERR · UHD-LL | PSNR: 27.57 -> 27.25 | Hurt |
| Add zero-frequency regularization (ablation) | ERR · UHD-LL | PSNR: 26.73 -> 27.57 | Helped |
| Add low-frequency regularization (ablation) | ERR · UHD-LL | PSNR: 26.57 -> 27.57 | Helped |
| Add high-frequency regularization (ablation) | ERR · UHD-LL | PSNR: 26.80 -> 27.57 | Helped |
| Add the zero-frequency enhancer (ZFE) (ablation) | ERR · UHD-LL | PSNR: 27.01 -> 27.57 | Helped |
Surprising findings
- Swapping only the zero-frequency component shows that an exchanged input with a degraded non-zero spectrum can score higher PSNR than an exchanged GT, implying the global DC term is critical early in restoration.
Limitations
- Performance remains unsatisfactory in extremely low-light scenarios, as shown in the failure case figure.
- Frequency cutoff k is a hyperparameter whose best value (k=64) was selected by ablation; performance drops at other values.
Experimental setup
Datasets: UHD-LL 4K-Rain13k UHD-Haze UHD-Blur
Benchmarks: UHD-LL 4K-Rain13k UHD-Haze UHD-Blur
Metrics: PSNR SSIM LPIPS parameters inference time
Baselines: IFT SNR-Aware Uformer Restormer DiffLL LLFormer UHDFour UHDFormer Wave-Mamba UDR-Mixer UDR-S2Former DehazeFormer
Builds on: UHDFour UHDformer GR-KAN Vision Mamba Restormer
Closest papers by meaning
- UHD-processer: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts (CVPR 2025 · similarity 0.91)
- Latent Harmony: Synergistic Unified UHD Image Restoration via Latent Space Regularization and Controllable Refinement (NeurIPS 2025 · similarity 0.89)
- DreamUHD: Frequency Enhanced Variational Autoencoder for Ultra-High-Definition Image Restoration (AAAI 2025 · similarity 0.89)
- Correlation Matching Transformation Transformers for UHD Image Restoration (AAAI 2024 · similarity 0.88)
- UltraHR-100K: Enhancing UHR Image Synthesis with A Large-Scale High-Quality Dataset (NeurIPS 2025 · similarity 0.87)
- ReFIR: Grounding Large Restoration Models with Retrieval Augmentation (NeurIPS 2024 · similarity 0.87)
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.