New AI Framework Tackles Multiple Image Flaws at Once with Adaptive Routing

Image restoration gets considerably harder when photos are damaged in multiple ways simultaneously—a problem most existing systems sidestep by focusing on one type of flaw at a time. A new approach called CART (Composite-Adaptive Routing and Task Conditioning) tackles this real-world challenge by dynamically deciding both where to concentrate processing effort and which degradation patterns to target.

The Problem with Mixed Damage#

Real photographs often suffer from several overlapping issues—blur, noise, compression artifacts, and more—all at once. Traditional restoration tools use the same amount of computation everywhere in the image and treat degradation as a single category. This blanket approach wastes resources on areas that are already clean and misses the nuance of handling multiple simultaneous problems.

How CART Works#

The system employs two key decision-making layers. A spatial mixer examines each patch of the image and routes it to specialized experts based on how badly it needs restoration, concentrating computing power where damage is worst. Separately, a channel mixer guides the image through degradation-specific pathways, informed by what types of flaws are actually present. This dual routing uses multi-hot classification—identifying multiple degradation types at once rather than forcing a single label—to condition the restoration process.

Performance Gains#

In experiments, CART achieved state-of-the-art results on the CDD-11 benchmark, which tests composite degradation restoration. It also outperformed prior methods on conventional multi-task benchmarks covering 3 and 5 different restoration challenges. By adapting both where and how it restores, the approach sidesteps inefficiency while handling real-world complexity that most all-in-one methods ignore.