Research map / Image generation / Image restoration
Navigating Image Restoration with VAR's Distribution Alignment Prior
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.
Uses multi-scale latent representations from a pretrained VAR (next-scale prediction) as distribution-alignment priors, injected via degradation-aware and adaptive feature transformation modules, to build VarFormer, an all-in-one image restoration model.
Problem
Multi-task (all-in-one) image restoration struggles with diverse unpredictable degradations; existing methods rely on the degraded input and neglect the clean-image distribution prior, and generalization to unseen degradations is limited.
Approach
Observes that VAR's next-scale prediction progressively aligns degraded-image latent representations with the clean-image distribution, and treats VAR's multi-scale latents as restoration priors. A two-stage design freezes VAR with an adapter and feature-matching loss (stage 1), then uses degradation-aware enhancement (scale weighting), adaptive feature transformation, and adaptive mix-up skips to
Key results
- On six-task all-in-one restoration, VarFormer reports 31.33 PSNR on 5-set deraining, 30.99 on GoPro deblurring, 25.13 on low-light enhancement, and 32.96 on dehazing, outperforming listed universal methods like DiffUIR.
- On SIDD real denoising, VarFormer achieves 40.13 PSNR / 0.978 SSIM, slightly above DiffUIR (40.11 / 0.976).
- In zero-shot under-display camera restoration on TOLED, VarFormer reaches 30.61 PSNR vs 29.55 for DiffUIR.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use VarFormer instead of the strongest listed universal baseline | All-in-one image restoration · Low-light image enhancement | PSNR: 25.02 -> 25.13 | Helped |
| Use VarFormer instead of the strongest listed universal baseline | All-in-one image restoration · Dehazing | PSNR: 32.74 -> 32.96 | Helped |
| Use VarFormer instead of the strongest listed universal baseline | Gaussian denoising · BSD68, σ=15 | PSNR: 34.05 -> 34.11 | Helped |
| Use VarFormer instead of the strongest listed universal baseline | Gaussian denoising · Urban100, σ=50 | PSNR: 28.45 -> 28.96 | Helped |
| Use VarFormer instead of the strongest listed universal baseline | Zero-shot UDC restoration · TOLED | PSNR: 29.55 -> 30.61 | Helped |
| Use VarFormer instead of the strongest listed baseline | Zero-shot UDC restoration · POLED | PSNR: 16.71 -> 16.63 | Hurt |
| Add DAE (ablation) | VarFormer · Six-task average | PSNR: 29.37 -> 29.66 | Helped |
| Add AFT (ablation) | VarFormer · Six-task average | PSNR: 29.51 -> 29.66 | Helped |
| Add Adapter (ablation) | VarFormer · Six-task average | PSNR: 29.29 -> 29.66 | Helped |
| Add adaptive mix-up skip (ablation) | VarFormer · Six-task average | PSNR: 28.98 -> 29.66 | Helped |
Surprising findings
- Replacing lower-scale VAR features removes global degradations (low light, haze) while higher-scale features remove local ones (noise, rain).
Limitations
- Ablation reports only average over six tasks and a single weight configuration
- Efficiency/training-cost claims are not quantified in the text
- Compute hardware not specified
Experimental setup
Datasets: Rain13K Rain100L Rain100H Test100 Test1200 Test2800 RESIDE LOL GoPro BSD400 WED BSD68
Benchmarks: Rain100L Rain100H Test100 Test1200 Test2800 RESIDE LOL GoPro BSD68 Urban100 Kodak24 SIDD
Metrics: PSNR SSIM LPIPS
Baselines: SwinIR MIRNet Restormer MAXIM RDDM AirNet Painter IDR DA-CLIP Prompt-IR DiffUIR NAFNet
Builds on: VAR AWRCP Restormer DiffUIR
Closest papers by meaning
- Distilling Semantic Priors from SAM to Efficient Image Restoration Models (CVPR 2024 · similarity 0.86)
- UniRestore: Unified Perceptual and Task-Oriented Image Restoration Model Using Diffusion Prior (CVPR 2025 · similarity 0.85)
- URWKV: Unified RWKV Model with Multi-state Perspective for Low-light Image Restoration (CVPR 2025 · similarity 0.83)
- Posterior-Mean Rectified Flow: Towards Minimum MSE Photo-Realistic Image Restoration (ICLR 2025 · similarity 0.82)
- Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.82)
- SP$^3$: Spherical Priors for Plug-and-Play Restoration (NeurIPS 2026 · similarity 0.81)
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.