liweileev / fet-gan Goto Github PK

Hi , I want to use model to train grayscale fonts images , how can I do it
If I change input_nc to 1 , the following error will occur
RuntimeError: output with shape [1, 128, 128] doesn't match the broadcast shape [3, 128, 128]

Is it necessary to change more codes to achieve this effect which I purposed .
What should I do to achieve training grayscale images?

Thanks in advance !

Thanks in advance.

I use Fonts100 pre-trained network and it works fine with characters in the font 100 datasets.
Following figure-9 in the paper, I try to experiment with new characters.
I use PIL to draw text and have the following result.

I found that the network is very sensitive to the input image. Even changing the font size would make big difference. Can you please share scripts that can create input images of new characters?

Thanks in advance !

Hi @liweileev ，thank you for your paper code. I get some questions about Fonts100's test set. Is the style of the source of this test set randomly selected?
Thank you!

Hi, thank you for your paper code!

I would like to test the pretrained networks on custom text effects and fonts. So far I haven't been getting the desired results with new styles. How could I finetune the pretrained models ?

HI, after I trained 20 epochs followed by your code. The reconstruction result is good but it can not transfer any style from reference images

I'm running code and style shifting doesn't have any effect, the refactoring works pretty well do you know why

@liweileev Nice to see what you do, your work is really interesting. I have some doubts about the generation of your dataset, Could you please provide me with the script that generates your data.
Thank you very much!