Figure 2: Overview of the proposed RestoreMore framework. A pretrained restoration model is
preserved as a frozen capability anchor, and lightweight residual expansion modules are inserted at
selected feature stages. Inside each expansion module, a degradation descriptor is first constructed
from spatial and frequency statistics. Top-M capability selection is then performed over accumu-
lated capability prototypes and routers. The selected routers further retrieve complementary experts
from the expert-key bank through query–key matching. Frozen historical components and trainable
current-task components are indicated by snowflake and flame icons, respectively.