Hu Gao

GraLoD, a plug-and-play framework that treats restoration scale as a spatially varying and stage-dependent continuous variable, is proposed and minimal-sufficient footprint calibration (MSFC) together with structure-aware regularization (SAR) is introduced to encourage restoration-effective and spatially coherent LOD assignments.

Existing image restoration models can now learn new tasks without sacrificing previously acquired skills, thanks to a novel capability-oriented routing mechanism.

Causal-AgentIR enables image restoration agents to dynamically evolve their knowledge, leading to more effective handling of diverse degradation scenarios.