Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration
Researchers have introduced Continuous Expert Assembly (CEA), a novel framework designed for all-in-one image restoration tasks. Addressing the challenge of unknown, spatially non-uniform, and compositional real-world image degradations, CEA adapts a single set of weights to diverse local corruption patterns without requiring test-time degradation labels. Unlike existing methods that rely on global prompts or static expert routing, which can bottleneck localized evidence or produce homogeneous updates, CEA employs a lightweight Cross-Attention Hyper-Adapter. This component probes intermediate spatial features to synthesize instance-conditioned low-rank routing bases and residual directions. Each spatial token then assembles its own residual update through dense signed dot-product affinities, avoiding external prompts and discrete selection mechanisms. Experimental results on AIO-3, AIO-5, and CDD-11 datasets demonstrate that CEA significantly improves average restoration quality compared to strong baselines, particularly for spatially varying and compositional degradations. The method maintains favorable efficiency in terms of parameters, FLOPs, and runtime, offering a transparent and effective solution for complex image restoration scenarios.
Wire timeline
Continuous Expert Assembly: Instance-Conditioned Low-Rank Residuals for All-in-One Image Restoration
Researchers have introduced Continuous Expert Assembly (CEA), a novel framework designed for all-in-one image restoration tasks. Addressing the challenge of unknown, spatially non-uniform, and compositional real-world image degradations, CEA adapts a single set of weights to diverse local corruption patterns without requiring test-time degradation labels. Unlike existing methods that rely on global prompts or static expert routing, which can bottleneck localized evidence or produce homogeneous updates, CEA employs a lightweight Cross-Attention Hyper-Adapter. This component probes intermediate spatial features to synthesize instance-conditioned low-rank routing bases and residual directions. Each spatial token then assembles its own residual update through dense signed dot-product affinities, avoiding external prompts and discrete selection mechanisms. Experimental results on AIO-3, AIO-5, and CDD-11 datasets demonstrate that CEA significantly improves average restoration quality compared to strong baselines, particularly for spatially varying and compositional degradations. The method maintains favorable efficiency in terms of parameters, FLOPs, and runtime, offering a transparent and effective solution for complex image restoration scenarios.
cs.AI updates on arXiv.org