Research map / Image generation / Image restoration
UHD-processer: Unified UHD Image Restoration with Progressive Frequency Learning and Degradation-aware Prompts
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 UHD-Processor, a VAE-based latent-space all-in-one restoration framework for UHD images with frequency-progressive curriculum training and low-rank degradation-aware and frequency-selection prompts.
Problem
All-in-one image restoration for ultra-high-definition (UHD) images is computationally expensive, since existing AIO methods rely on heavy restoration backbones and redundant degradation-aware prompt branches that do not scale to 4K inputs.
Approach
Moves restoration into a compact VAE latent space with a simple latent restoration network (Cubic-Mixer), trains sub-networks progressively from low to high wavelet frequency bands (curriculum), and uses low-rank, scale-invariant degradation-aware prompts in the encoder and frequency-selection prompts in the decoder.
Key results
- On four degradations, average PSNR 29.23 dB with 1.6M params and 4G FLOPs, beating the second-best HAIR by 1.93 dB.
- On six degradations, average PSNR 28.67 dB, improving by 1.7 dB over Histoformer with ~9.6% of its parameters.
- Frequency-progressive learning improves PSNR by 1.99 dB over direct mixed training (27.24 to 29.23 in ablation).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FDPL instead of mixed-data training (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 27.24 -> 29.23 | Helped |
| use SL1 order: noise→blur→low-light→haze (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 27.24 -> 27.08 | Hurt |
| use SL2 order: blur→noise→haze→low-light (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 27.24 -> 28.12 | Helped |
| add DALP (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 27.86 -> 29.23 | Helped |
| add DDP (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 28.49 -> 29.23 | Helped |
| add DTP (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 28.34 -> 29.23 | Helped |
| add DSR (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 28.84 -> 29.23 | Helped |
| add MCKG (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 28.72 -> 29.23 | Helped |
| add DFSP (ablation) | UHD-Processor · UHD all-in-one, four degradations | PSNR: 28.21 -> 29.23 | Helped |
| use UHD-Processor instead of Histoformer | UHD all-in-one restoration · Average, six degradations | PSNR: 26.59 -> 28.67 | Helped |
Surprising findings
- Sequential task-based learning orders (SL1) can degrade performance, while the optimal order depends on data, so a fixed task order is hard to define.
Limitations
- Relies on a two-stage pipeline with a pre-trained VAE and a separately trained degradation classifier.
- Benchmark-based evaluation on synthetic or curated UHD datasets; UHD-noise is constructed synthetically.
Experimental setup
Datasets: UHD-LL UHD-blur UHD-haze UHD-rain UHD-snow UHD-noise (constructed from 4K images in [42])
Benchmarks: UHD-LL UHD-blur UHD-haze UHD-noise UHD-rain UHD-snow
Metrics: PSNR SSIM LPIPS FLOPs parameters
Baselines: AirNet IDR PromptIR CAPTNet NDR-Restore Gridformer DiffUIR-L Histoformer adaIR HAIR Restormer UHDformer
Builds on: PromptIR Curriculum learning Cubic-Mixer Restormer LoRA
Closest papers by meaning
- DreamUHD: Frequency Enhanced Variational Autoencoder for Ultra-High-Definition Image Restoration (AAAI 2025 · similarity 0.91)
- From Zero to Detail: Deconstructing Ultra-High-Definition Image Restoration from Progressive Spectral Perspective (CVPR 2025 · similarity 0.91)
- Correlation Matching Transformation Transformers for UHD Image Restoration (AAAI 2024 · similarity 0.89)
- Latent Harmony: Synergistic Unified UHD Image Restoration via Latent Space Regularization and Controllable Refinement (NeurIPS 2025 · similarity 0.89)
- UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.86)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 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.