why not one model for restore all degrad?

[foocker]

为何是一个架构，不同低质量模式，训练不同模型权重，而不是一个模型，所有低质量模式？

[gy65896]

Yes, our model solves multiple problems with one model, so you can implement this using a model trained on our dataset (CDD-11). Training on other benchmarks is just to verify the generalization of the model.

[foocker]

Yes, our model solves multiple problems with one model, so you can implement this using a model trained on our dataset (CDD-11). Training on other benchmarks is just to verify the generalization of the model.

o, i see. another question is: the inference time of 2k*2k input is ?

[foocker]

Yes, our model solves multiple problems with one model, so you can implement this using a model trained on our dataset (CDD-11). Training on other benchmarks is just to verify the generalization of the model.

我简单测试了一下，问题1.图像太大，resize scale 太大，attention的计算显存要求巨高，比如4k图要80G左右；问题2.输入图像形状不同，会出bug，整除之类的bug；问题3.官方给的例子图，出来的细节很差。和模型大小有关？
整体推理时间，还是不错的。