Thanks for your attention to our work. The pre-trained model named OneRestore trained on our CDD-11 for Real Scenes (onerestore_real.tar) was also trained using CDD-11. It doesn't require a special dataset or a new loss function. It follows the same training settings described in Sec.5.1 of the paper.

Thanks for your reply, but I have a further question. What is the difference between [onerestore_cdd-11.tar](https://1drv.ms/u/s!As3rCDROnrbLgqpmWkGBku6oj33efg?e=7yUGfN) and [onerestore_real.tar](https://1drv.ms/u/s!As3rCDROnrbLgqpi-iJOyN6OSYqiaA?e=QFfMeL). It seems that they are both trained on CDD11, with the same loss function, epochs, learning-rate, optimizer, etc.

Same question here.

Thank you for your question. Specifically, both models are trained under the same configuration. The difference is that for the first model, we selected the result with the best indicator for experimental comparison, while the other one is the model of the last batch.

So, onerestore_cdd-11.tar is the best parameter for CDD11-testset while onerestore_real.tar is the parameter of the last epoch. But why the parameter in the last epoch can be for Real Scenes?

In practice, this is not different, but since we used different model weights for synthetic and real data in the paper, we publish both weights to ensure that the implementation results match our paper.

Could you explain why you choose to use the parameters of the final epoch for real data?

There is no special reason, we publish this just to allow researchers to reproduce the same results in our paper.

Well, let me describe my query in Chinese...

根据您的解释，OneRestore并没有使用额外的训练方式、新增的退化或数据集，仅仅使用最后一轮的的权重去处理真实退化场景的图像。

那么，onerestore_cdd-11.tar不能处理真实退化图像吗？如果可以，为什么文章中要采用onerestore_real.tar去处理真实图像？如果不行，为什么onerestore_cdd-11.tar不能处理真实退化图像？

I guess the authors just did this in implementation but not with a clear purpose. It may be okay to use the same weights for real scenes. Have you tested the onerestore_cdd-11.tar for real scenes?

Well, let me describe my query in Chinese...

根据您的解释，OneRestore并没有使用额外的训练方式、新增的退化或数据集，仅仅使用最后一轮的的权重去处理真实退化场景的图像。

那么，onerestore_cdd-11.tar不能处理真实退化图像吗？如果可以，为什么文章中要采用onerestore_real.tar去处理真实图像？如果不行，为什么onerestore_cdd-11.tar不能处理真实退化图像？

You are right. As I mentioned before, these two checkpoints have no significant difference in model training. It does not mean that onerestore_cdd-11.tar can not process the real image. We just want researchers can produce the same results as our paper. Please don't dwell on this.

I guess the authors just did this in implementation but not with a clear purpose. It may be okay to use the same weights for real scenes. Have you tested the onerestore_cdd-11.tar for real scenes?

Well, let me describe my query in Chinese...
根据您的解释，OneRestore并没有使用额外的训练方式、新增的退化或数据集，仅仅使用最后一轮的的权重去处理真实退化场景的图像。
那么，onerestore_cdd-11.tar不能处理真实退化图像吗？如果可以，为什么文章中要采用onerestore_real.tar去处理真实图像？如果不行，为什么onerestore_cdd-11.tar不能处理真实退化图像？

This implementation confuses me a lot. The authors just do this but don't clearly state the reason for using this setting in their paper or repository description, which makes me doubt these results of this paper.

We may never know the authors' intention in conducting the test for real data in such way.

If you have any doubts, you can just do the experiment. I have made all the checkpoints public.

Dear authors of OneRestore,

It is the author's responsibility to fully explain the setting of your paper and the intention behind it.

And your "strongly worded" feedback will cause more researchers to misunderstand your work. You keep ducking this issue for no reasons.

Maybe you just conduct the wrong test, or maybe there is a scientific research topic behind this: perhaps extended training time, appropriate parameter selection or stacking can increase the generalization of image restoration models under complicated and heterogeneous real degradations?

As readers, we cannot know why, and it is difficult to fully replicate your experiments. We have our own experiments. We desire an explanation, even if it is due to your errors or omissions. Explanation is preferable to evading the situation.

Perhaps you could wait until you are emotionally stable before answering the questions above.