Under the SEAL framework's 100 representative degradation cases, several Real-SR methods produce output worse than a small FSRCNN baseline on more than half the cases. Among MSE-based methods, SRResNet (AR: 0.00), DASR (AR: 0.00), and RD-SR (AR: 0.08) all fail. Among GAN-based methods, ESRGAN (AR: 0.00) and DASR (AR: 0.02) also fail. The condition is diverse real-world degradations spanning blur, noise, resize, and JPEG compression, and the failure is checked across all 100 representative cases.
Evidence
correlational
Key metric
AR: 0.00 (SRResNet), 0.00 (DASR), 0.08 (RD-SR) for MSE-based; 0.00 (ESRGAN), 0.02 (DASR) for GAN-based, out of 100 representative degradation cases
Caveat
The acceptance line is a small FSRCNN (0.4M params) fine-tuned per degradation case; a different reference line could shift the threshold.