UniRestorer++: Scalable Universal Image Restoration with Multi-Granularity Degradation Modeling
The authors report stronger scaling gains on unfamiliar and real-world image damage than standard mixture-of-experts scaling.
UniRestorer++ adds specialized experts as it divides image degradations into finer groups. Its routing chooses both an expert path and a specialization level to balance precision with robustness, with region-aware routing for damage that varies across an image. The researchers also introduce CDD-50, a benchmark covering 50 categories with up to five combined degradations. Across the tested model sizes, they report approximately log-linear gains in PSNR, an image reconstruction quality metric.
HF Daily Papers' note
UniRestorer++ adds specialized experts as it divides image degradations into finer groups. Its routing chooses both an expert path and a specialization level to balance precision with robustness, with region-aware routing for damage that varies across an image. The researchers also introduce CDD-50, a benchmark covering 50 categories with up to five combined degradations. Across the tested model sizes, they report approximately log-linear gains in PSNR, an image reconstruction quality metric.
HF Daily Papers' note
score 5