Training Datasets Used Question

[hezhiyang2000]

Dear author,

I hope this message finds you well. While engaging with your work and attempting to replicate the experiments based on the realsr_swinunet_realesrgan256.yaml configuration file provided, I came across a detail regarding the usage of datasets during training. The configuration file lists two dataset paths as follows:

      txt_file_path: [
                      '/mnt/lustre/zsyue/database/ImageNet/train/image_path_all.txt', 
                      '/mnt/lustre/zsyue/database/FFHQ/files_txt/ffhq256.txt',
                     ] 

However, in the relevant section of your paper, while ImageNet dataset is mentioned as a resource for training, there isn't an explicit indication that the FFHQ dataset is also included. Consequently, I would like to seek clarification from you regarding which datasets were actually used in the model training and experimental results reported in your paper. Could you please clarify whether the outcomes presented in your research were derived solely from the ImageNet dataset or if they also incorporated data from the FFHQ dataset? Thank you!

[zsyOAOA]

In our initial version, we adopt only the ImageNet dataset. However, we found some previous work (BSRGAN, RealESRGAN) have added some face images to the training dataset for better performance on face images. Thus, we then also adopt such a strategy following them.

[hezhiyang2000]

Thank you for your prompt response and clarification. Based on your reply, I understand that the results presented in the ResShift paper indeed involve the use of the FFHQ dataset. Could you kindly provide specific details about how the FFHQ dataset was employed in your experiments, including the particular version and amount of data of the FFHQ dataset used and how it was integrated with the ImageNet dataset during the training process? Your detailed explanation is greatly appreciated once again.

[zsyOAOA]

The FFHQ dataset contains 70000 images with a resolution of 1024. I first resize them to the resolution of 256, and then directly add them into the training data.

[XLR-man]

@zsyOAOA Hi, I was wondering, can your model directly train on other low level vision tasks such as image enhancement task? Second, if the width and height of the dataset resolution are inconsistent, can it be trained directly? Or need to change the resolution?

[zsyOAOA]

I have released the code on other low-level vision tasks, e.g., face restoration, inpainting. @XLR-man

[649459021]

In our initial version, we adopt only the ImageNet dataset. However, we found some previous work (BSRGAN, RealESRGAN) have added some face images to the training dataset for better performance on face images. Thus, we then also adopt such a strategy following them.

Hi, Are you using aligned or unaligned face images?

[zsyOAOA]

I use the aligned face images from FFHQ. @649459021