Jingyun Liang

You could crop HR during the generation of LR, even it's testing set.

multiplied by 255 should not change the std. After converted to BGR, I guess the std should be [0.225, 0.224, 0.229]( the standard std is [0.229, 0.224, 0.225])

Sorry for the late reply. netG and netE means the commonly used Generator and the Generator_EMA (the interpolation between models of the last iteration and the current iteration. It leads...

\#Params: use ``` print({sum(map(lambda x: x.numel(), model.parameters())):,d}) ``` \# FLOPS: refer to https://github.com/JingyunLiang/SwinIR/blob/510b75617620b05f4e35f74dec8d30d59c721cf6/models/network_swinir.py#L862

The artifact comes from dividing the whole image into patches during testing (due to limited memory). I think retraining the model can alleviate the problem.

1, It's strange that your L1 loss trained model performs bad. I currently have no idea why it happens. Is the loss normal? 2, If you meet I/O bottleneck, preparing...

No. We haven't trained yet. We just followed the common setting in JPEG_CAR. React with thumbs up here if you want such a model. I will train and release one...

Great! I will train it today. Do you want four separate models like [006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth](https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/006_CAR_DFWB_s126w7_SwinIR-M_jpeg10.pth) [006_CAR_DFWB_s126w7_SwinIR-M_jpeg20.pth](https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/006_CAR_DFWB_s126w7_SwinIR-M_jpeg20.pth) [006_CAR_DFWB_s126w7_SwinIR-M_jpeg30.pth](https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/006_CAR_DFWB_s126w7_SwinIR-M_jpeg30.pth) [006_CAR_DFWB_s126w7_SwinIR-M_jpeg40.pth](https://github.com/JingyunLiang/SwinIR/releases/download/v0.0/006_CAR_DFWB_s126w7_SwinIR-M_jpeg40.pth) Or a joint model for JPEG10~40?

Yes, I agree. Training separate models have best performance. I will update here when the RGB models are released.

See results for Colored JPEG Compression Artifacts Removal in https://github.com/JingyunLiang/SwinIR#results and pretrained models at https://github.com/JingyunLiang/SwinIR/releases.