How to use higher resolution images for model training？？ #329

New Issue

claunia · 2026-01-29T21:46:54Z

claunia commented

2026-01-29 21:46:54 +00:00

Originally created by @BruceZhou95 on GitHub (Mar 29, 2023).

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

Originally created by @BruceZhou95 on GitHub (Mar 29, 2023). Hello, Hello The image input size of the original model input is 512*512. I would like to ask how to use a higher resolution image for training, such as 1024*1024. Which file's code needs to be modified. Thank you very much

claunia commented

2026-01-29 21:46:55 +00:00

@lightshadown commented on GitHub (Mar 30, 2023):

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

@lightshadown commented on GitHub (Mar 30, 2023): 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

claunia commented

2026-01-29 21:46:55 +00:00

@BruceZhou95 commented on GitHub (Mar 30, 2023):

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.

@BruceZhou95 commented on GitHub (Mar 30, 2023): > 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.

claunia commented

2026-01-29 21:46:55 +00:00

@YilanWang commented on GitHub (Apr 11, 2023):

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 .

@YilanWang commented on GitHub (Apr 11, 2023): 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 .

claunia commented

2026-01-29 21:46:56 +00:00

@BruceZhou95 commented on GitHub (Apr 12, 2023):

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.

@BruceZhou95 commented on GitHub (Apr 12, 2023): > 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.

claunia commented

2026-01-29 21:46:56 +00:00

@marshalleq commented on GitHub (Jul 16, 2023):

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.

@marshalleq commented on GitHub (Jul 16, 2023): 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.

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: TencentARC/GFPGAN#329