Research map / Image generation / Image restoration

Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration

NeurIPS 2026 · Image generation › Image restoration · approach: Efficient restoration architectures · 0 citations (OpenAlex)

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 Continuous Expert Assembly (CEA), which uses a cross-attention hyper-adapter to generate instance-conditioned low-rank routing bases and residual directions and assembles per-token residual updates via dense signed dot-product affinities for all-in-one image restoration.

Problem

All-in-one image restoration must handle unknown, spatially non-uniform and compositional degradations with a single set of weights; global prompts/descriptors bottleneck localized evidence, while static expert pools and sparse Top-K routing give homogeneous or unstable updates.

Approach

Instead of routing among a fixed expert bank, a cross-attention hyper-adapter (learnable queries probing condensed spatial features) generates image-conditioned low-rank factors; each token's residual is a softmax-free signed dot-product assembly over these rank-wise components, injected into attention Q/K projections.

Key results

• On CDD-11, CEA-IR-S reaches 29.79/0.886 average vs. 29.05/0.881 for MoCE-IR-S with 10M parameters.
• On AIO-5 heavy setting, CEA-IR reaches 31.13/0.922 average vs. 30.58/0.919 for MoCE-IR.
• Dynamic factors with dense assembly improve over Top-2 softmax routing on CDD-11 by 0.59 dB (29.01 to 29.79 avg PSNR).

Surprising findings

• Static (input-independent) low-rank factors with dense assembly remain close to the backbone-only result (29.01 vs 29.00 avg PSNR on CDD-11).

Limitations

• Evaluated only on standard restoration benchmarks; generalization to other restoration settings is untested.
• Implemented only on a U-shaped Transformer backbone modulating attention Q/K projections.
• Gains over MoCE-IR-S are modest (e.g., +0.32 dB on AIO-3), and the bootstrap analysis measures test-set paired significance rather than training-seed variance.

Experimental setup

Datasets: AIO-3 training mixture (dehazing, deraining, denoising) AIO-5 training mixture CDD-11

Benchmarks: SOTS Rain100L BSD68 GoPro LOLv1 CDD-11 UIEB

Metrics: PSNR SSIM FLOPs runtime parameters

Baselines: MoCE-IR MoCE-IR-S OneRestore PromptIR AirNet LoRA-IR Res-HAIR VLU-Net IDR MIRAGE-S

Builds on: MoCE-IR PromptIR LoRA Restormer

Closest papers by meaning

• Complexity Experts are Task-Discriminative Learners for Any Image Restoration (CVPR 2025 · similarity 0.90)
• Learning Where and What to Restore for Composite Image Restoration (NeurIPS 2026 · similarity 0.89)
• Visual-Instructed Degradation Diffusion for All-in-One Image Restoration (CVPR 2025 · similarity 0.88)
• Parameter Efficient Adaptation for Image Restoration with Heterogeneous Mixture-of-Experts (NeurIPS 2024 · similarity 0.88)
• Residual-Conditioned Optimal Transport: Towards Structure-Preserving Unpaired and Paired Image Restoration (ICML 2024 · similarity 0.87)
• Debiased All-in-one Image Restoration with Task Uncertainty Regularization (AAAI 2025 · similarity 0.87)