Research map / Image generation / Image restoration

Learning Where and What to Restore for Composite Image Restoration

NeurIPS 2026 · Image generation › Image restoration · approach: Efficient restoration architectures

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 CART, a composite image restoration U-Net with a patch-level spatial mixer (graded-capacity experts) and an image-level channel mixer routed by multi-hot degradation classification, reaching SOTA on CDD-11 and conventional 3D/5D all-in-one benchmarks.

Problem

Composite image restoration, where multiple co-occurring degradations (low-light, haze, rain, snow) must be restored in one input without known degradation labels; existing all-in-one methods use uniform spatial computation and single-label task conditioning.

Approach

Combines patch-wise content-adaptive routing to graded-capacity attention experts (where to compute) with image-level channel-expert routing driven by a multi-hot degradation prediction from a bottleneck task feature (what degradation cues to use), trained with BCE multi-hot supervision and load-balancing losses.

Key results

• CART attains average PSNR 29.95 dB on CDD-11, +0.90 dB over MoCE-IR (29.05), ranking best on 10 of 11 conditions in PSNR.
• On the 3D all-in-one benchmark CART reaches 33.10 dB average PSNR vs 32.73 for MoCE-IR; on 5D it reaches 30.87 vs 30.58.
• CART uses 19.47 GFLOPs (26% less than MoCE-IR) and runs at 32.11 ms, but has 32.58M parameters vs MoCE-IR's lower count.

Surprising findings

• Removing spatial routing costs only 0.05 dB on CDD-11 average, while removing channel routing costs 0.39 dB.
• Removing both routers drops average PSNR by 0.52 dB, slightly above the sum of individual effects.

Limitations

• MoE design stores all experts, giving higher parameter count and peak memory than MoCE-IR.
• Evaluation limited to synthetic benchmarks; real-world perceptual studies left to future work.
• Gains on well-saturated tasks such as dehazing and deblurring are small (on par with best prior).

Experimental setup

Datasets: CDD-11 BSD400 WED BSD68 Rain100L SOTS GoPro LOL-v1

Benchmarks: CDD-11 BSD68 Rain100L SOTS GoPro LOL-v1

Metrics: PSNR SSIM FLOPs runtime

Baselines: AirNet PromptIR WGWSNet WeatherDiff OneRestore MoCE-IR MPRNet Gridformer Art-PromptIR DA-CLIP UniProcessor NAFNet

Builds on: MoCE-IR Restormer MDTA OneRestore PromptIR

Closest papers by meaning

• Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration (NeurIPS 2026 · similarity 0.89)
• Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.88)
• Test-Time Degradation Adaptation for Open-Set Image Restoration (ICML 2024 · similarity 0.87)
• Degradation-Aware Feature Perturbation for All-in-One Image Restoration (CVPR 2025 · similarity 0.87)
• Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.87)
• UP-Restorer: When Unrolling Meets Prompts for Unified Image Restoration (AAAI 2025 · similarity 0.87)