Issues about Image Size

[Z-yiwei]

Dear author, thanks for your excellent work! After carefully reading your paper and your code, I see that the training is based on 64-256 SR task. But it seems that you use the pretrained weights (VQGAN encoder/decoder and SwinUnetModel) from 64-512 SR task to apply in 128-512 SR Task. How does it work by applying the same weights for different tasks? Or do I misunderstood your work? Thank you!

[zsyOAOA]

You may misunderstand some details. I do not apply the pretrained weight for the task 65-512 SR. After training, our model is able to deal with any X4 SR task on any resolution. @Z-yiwei

[Z-yiwei]

Thank you very much! And sorry for my typo ( I originally mean 64-256 SR task and 128-512 task).

[LearningHx]

Sorry to be bother. I successfully completed the 64-256 SR task by using your model, but when I tested it on the 128-512 SR task, I found a significant decrease in performance. Do we need to make some adjustments to the parameters when the resolution is not 64?

[zsyOAOA]

You may need to increase the chopping size, please carefully read the inference code. Our model can be used for any resolution, such as the real examples shown in this repo. @LearningHx

[ccountingstars]

You may need to increase the chopping size, please carefully read the inference code. Our model can be used for any resolution, such as the real examples shown in this repo. @LearningHx您可能需要增加斩波大小，请仔细阅读推理代码。我们的模型可用于任何分辨率，例如此存储库中显示的真实示例。@LearningHx Thank you for checking my question in your busy schedule, running your model directly doesn't seem to be able to do anything with images of size 32*32 and below, we tried many times but couldn't solve the reported error, what modification to the code is required.