IC-768MMRealSR exhibits the most consistent performance across degradation cases while SwinIR achieves the highest performance on cases it handles well, among GAN-based Real-SR methods

Wenlong Zhang, Xiaohui Li, Xiangyu Chen, Xiaoyun Zhang, Yu Qiao, Xiao-Ming Wu, Chao Dong

SourceSEAL: A Framework for Systematic Evaluation of Real-World Super-Resolution

Among GAN-based Real-SR methods evaluated under SEAL, MMRealSR achieves the lowest interquartile range of RPR (RPR_i: 0.08), indicating the most uniform relative improvement across all 100 degradation cases. SwinIR, by contrast, has a higher RPR_i (0.24) but achieves a higher average RPR on acceptable cases (RPR_a: 0.71 vs 0.57 for MMRealSR), meaning it performs better on the cases it can handle. Both achieve high acceptance rates (MMRealSR AR: 0.80, SwinIR AR: 0.81), but they differ in the shape of their performance distribution.

Evidence
correlational
Key metric
RPR_i: 0.08 (MMRealSR) vs 0.24 (SwinIR); RPR_a: 0.71 (SwinIR) vs 0.57 (MMRealSR); AR: 0.80 (MMRealSR) vs 0.81 (SwinIR)
Caveat
The ranking depends on which metric is prioritized in the coarse-to-fine protocol; if RPR_a is used as the first fine-grained metric, SwinIR ranks first instead of MMRealSR.
Model
MMRealSR, SwinIR
Related work
BSRGAN [context]
Related findings
IC-766, IC-767
Extraction
automatic-extraction