I have a question ablout input_img size. about lffont HOT 2 CLOSED

I don't know how you define "qualitative" exactly, but I presume that you want to generate glyphs with "better" visual quality, e.g., fewer artifacts.

Also, I don't know how you define the input size. There can be two viewpoints:

Case 1. Train with 128 x 128, test with 256 x 256
Generally, it will not work. I presume that your question means case 2.

Case 2. Train with 256 x 256, test with 256 x 256
We cannot guarantee anything for this.
However, I think it does not bring meaningful advantages despite the memory and computation consumptions.
Note that by resizing the input size twice (4 times in pixel level), you will consume more than 4 times the memory and computation resources.

To sum up, I don't think resizing the input size will not be helpful for your case.
However, in some applications such as image classification, object detection, resizing large input size often brings higher performances (e.g., accuracies). So, I recommend you train your own model with 256 x 256 inputs.

from lffont.

OK, Thank you very much.

from lffont.