Research map / Image generation / Image restoration
UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration
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 Up-Restorer, an ADMM-unrolled, prompt-based score diffusion transformer that handles dehazing, deraining, and denoising in one model.
Problem
All-in-one image restoration must implicitly identify different degradation types and apply appropriate priors, but existing methods rely on task-specific heads, contrastive training, or multi-stage training, and diffusion restorers lack degradation-aware regularization.
Approach
Unroll the MAP-based ADMM solver for a unified degradation model (I = J·t + A·(1−t) + N) under a score-based diffusion model, where learned prompts generated per image update transmission, atmospheric light, and noise/rain maps at each stage; learnable queries are added to the diffusion transformer attention.
Key results
- In the all-in-one setting, averaged over dehazing, deraining and denoising, Up-Restorer reaches 32.16 dB PSNR vs 32.06 dB for PromptIR (about 0.10 dB gain).
- On Rain100L single-task deraining, Up-Restorer reports 39.52 dB PSNR vs 37.04 dB for PromptIR.
- Up-Restorer totals about 28.0M parameters, with the prompt upgrade module adding about 8% and learnable queries about 2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · SOTS | PSNR: 30.58 -> 30.68 | Helped |
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · Rain100L | PSNR: 36.37 -> 36.74 | Helped |
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · BSD68 σ=15 | PSNR: 33.98 -> 33.99 | Helped |
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · BSD68 σ=25 | PSNR: 31.31 -> 31.33 | Helped |
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · BSD68 σ=50 | PSNR: 28.06 -> 28.07 | Helped |
| Use Up-Restorer instead of PromptIR | all-in-one image restoration · Average across tasks | PSNR: 32.06 -> 32.16 | Helped |
| Use Up-Restorer instead of PromptIR | single-task restoration · SOTS | PSNR: 31.31 -> 31.12 | Hurt |
| Use Up-Restorer instead of PromptIR | single-task restoration · Rain100L | PSNR: 37.04 -> 39.52 | Helped |
| Add learnable queries, Prompt Block, and physics variables (ablation) | Up-Restorer · Rain100L | PSNR: 34.72 -> 36.74 | Helped |
| Add learnable queries, Prompt Block, and physics variables (ablation) | Up-Restorer · SOTS | PSNR: 29.23 -> 30.68 | Helped |
| Add physics variables (ablation) | Up-Restorer · Rain100L | PSNR: 34.64 -> 36.74 | Helped |
| Add physics variables (ablation) | Up-Restorer · SOTS | PSNR: 29.51 -> 30.68 | Helped |
Surprising findings
- Per-task results are mixed: on SOTS dehazing in the single-task table, Up-Restorer (31.12 dB) is slightly below PromptIR (31.31 dB) and Restormer (30.87 dB is lower, but PromptIR is higher).
Limitations
- Physics-variable interpretation (transmission map matching input) is shown only qualitatively via a visualization.
- Diffusion-based inference cost is not analyzed in the reported text.
Experimental setup
Datasets: BSD400 WED BSD68 Rain100L OTS RESIDE GoPro Snow100K RealSR TID2013 Koniq-10k
Benchmarks: SOTS Rain100L BSD68 GoPro Snow100K RealSR
Metrics: PSNR SSIM LPIPS HFD-IQA Hyper-IQA
Baselines: PromptIR AirNet WeatherDiffusion DA-CLIP IDR Restormer FDGAN MPRNet
Builds on: C2F-DFT ADMM score-based diffusion model PromptIR
Closest papers by meaning
- Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.89)
- GenDeg: Diffusion-based Degradation Synthesis for Generalizable All-In-One Image Restoration (CVPR 2025 · similarity 0.88)
- Scaling Up to Excellence: Practicing Model Scaling for Photo-Realistic Image Restoration In the Wild (CVPR 2024 · similarity 0.88)
- Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 2025 · similarity 0.88)
- Learning Diffusion Texture Priors for Image Restoration (CVPR 2024 · similarity 0.88)
- Prompt-tuning Latent Diffusion Models for Inverse Problems (ICML 2024 · similarity 0.88)
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.