How to use higher resolution images for model training？？

[BruceZhou95]

Hello, Hello
The image input size of the original model input is 512512. I would like to ask how to use a higher resolution image for training,
such as 10241024. Which file's code needs to be modified.
Thank you very much

[lightshadown]

i dont think is that easy, i think the limitation is the ammount of data generated when running the model, it creates too much and the GPU might get overflow, i have had issues when using a similar model from XINTAO/ESRGAN for super resolution on collab, if the file/resolution is too big, the gpu simply fails, so you might get better resoults if you take part of the generated data and create a tmp file to be save on disk but not sure how to modify the program, just brain storming here

[BruceZhou95]

i dont think is that easy, i think the limitation is the ammount of data generated when running the model, it creates too much and the GPU might get overflow, i have had issues when using a similar model from XINTAO/ESRGAN for super resolution on collab, if the file/resolution is too big, the gpu simply fails, so you might get better resoults if you take part of the generated data and create a tmp file to be save on disk but not sure how to modify the program, just brain storming here

Thank you for your reply. I have sufficient GPU resources and want to use my own 1024*1024 dataset for training, but I don't know which configuration files or parameters need to be changed.

[YilanWang]

And I have a problem, the ffhq dataset is 1024x1024, how to get the data resolution 512x512 in this paper? the author seems that he dosen't show this .

[BruceZhou95]

And I have a problem, the ffhq dataset is 1024x1024, how to get the data resolution 512x512 in this paper? the author seems that he dosen't show this .

At the beginning of the paper's experiment, the author wrote to resize the all images to 512*512.

[marshalleq]

I too would like the output to be 1024x1024. 512 is really for older GPU's now I think. I do everything 1024 because it does seem to get better results and I don't like losing quality. Surely this is possible to set as a flag then leave it up to the individual to ensure they have the right resources? Thanks.