Superresolution Transformer model PSNR isn't improving

Training data - 7313 satellite images
I am training on SwinIR model and the test results PSNR is stuck on around 20db. What could be the problem? Should I add more data? Originally SwinIR was trained on above 13,700 data natural images.