VITA-Group / ShapeMatchingGAN

[ICCV 2019, Oral] Controllable Artistic Text Style Transfer via Shape-Matching GAN
MIT License
423 stars 67 forks source link

question about different styles produce different deformation degrees #8

Closed Danee-wawawa closed 4 years ago

Danee-wawawa commented 4 years ago

Hi, I test my images on the pre-trained models you provied. But I found a bit strange in the test results. For example, in the case of the most deformation, ie scale=1, the test results of fire and water are as follows. image image Why is the effect of fire deformation so obvious but water has almost no deformation feeling?

williamyang1991 commented 4 years ago

I think there are two reasons. First, different stlyes have different deformations. You cannot expect every style to have obvious deformation effects as the fire. Second, your text is relatively large compared to the text used and trained in my paper. If the text is larrge, the scale of the deformation will consiquently look relatively small. Maybe you can train with your own data, with large style image size, large text image size and large blur kernel, to obtain obvious water-like deformations.

Danee-wawawa commented 4 years ago

Hi, my image is 320*320, just like yours, I just concat the results of the two words together. So as you say, if I want to obtain obvious water-like deformations, I can train with large blur kernel. The method you said ”large blur kernel” is to increase the kernel_size or increase the range 8-24?

williamyang1991 commented 4 years ago

69127328-0b348000-0ae5-11ea-9e17-9ccfa35641cd

image size != text size (font size) Large blur kerne means large standard deviation. So increasing the range 8-24 to 8-like 32 or 40.

Danee-wawawa commented 4 years ago

OK, Thank you~ So what is your font size and image size?

williamyang1991 commented 4 years ago

320*320 images, with text of 188.4pt size in Photoshop Please take the attached .psd for reference. text example.zip

Danee-wawawa commented 4 years ago

OK, thank you~^_^