How can we use this model to train on real low resolution images without super-resolution images? about dcls-sr

[Algolzw]

Although the ground truth kernels are required in training, we don't need them in testing. In other words, once the model is well-trained, we can directly apply it to any image, estimating the kernel and recovering a high-resolution version. So the inference is exactly the 'blind' super-resolution.

[javey-q]

Although the ground truth kernels are required in training, we don't need them in testing. In other words, once the model is well-trained, we can directly apply it to any image, estimating the kernel and recovering a high-resolution version. So the inference is exactly the 'blind' super-resolution.

作者您好，感谢您做出的优秀工作，关于以上讨论我还有些问题想请教一下：我看您的论文和代码中，在每次训练过程中，关于模糊核的类型和尺寸都是特定的，这样如果在不确定测试集的模糊核类型和大小的情况下，是否效果就不行了？
进一步地，为了解决上述问题，如果在训练过程中让模糊核的类型和尺寸都是随机的，这样您的方法还是否适用呢？
刚接触这个领域不久，有理解不对的还请谅解。

[Algolzw]

理论上确实如此，测试集模糊核不确定的话效果肯定会变差。不过我没有试过随机类型或尺寸模糊核，这个也许可以参考下Real-ESRGAN。