Continuous Expert Assembly for All-in-One Image Restoration
A new framework called Continuous Expert Assembly (CEA) is proposed for all-in-one image restoration, addressing unknown, spatially non-uniform, and compositional real-world degradation. CEA uses a Cross-Attention Hyper-Adapter to probe intermediate spatial features and generate instance-conditioned low-rank routing bases and residual directions, enabling token-wise dynamic parameterization. This approach overcomes limitations of global conditioning and static expert routing. The paper is published on arXiv with ID 2605.06127.
Key facts
- CEA is a token-wise dynamic parameterization framework for all-in-one image restoration.
- It uses a Cross-Attention Hyper-Adapter to synthesize instance-conditioned low-rank routing bases and residual directions.
- Real-world image degradation is often unknown, spatially non-uniform, and compositional.
- Existing methods use global prompts or predefined expert pools.
- Global conditioning can bottleneck localized degradation evidence.
- Static expert routing may produce homogeneous updates or rely on unstable sparse assignments.
- The paper is available on arXiv with ID 2605.06127.
- The announcement type is cross.
Entities
Institutions
- arXiv