search

Yijunmaverick / MultiTextureSynthesis

The source code of CVPR17 'Diversified Texture Synthesis with Feed-forward Networks'.
MIT License
55 stars 12 forks source link

about inference #3

Open suke27 opened 6 years ago

suke27 commented 6 years ago

Hi, during inference, I found the pad need be add for transfer, 1st could you tell why we need add pads. it seems hard to predict for different resolution (like w : 512, h:256)

Plz help to check, thank you!

suke27 commented 6 years ago

@Yijunmaverick , I found during inference, we may need add unit with input file, the unit tensor is 4 times smaller than input tensor. for padding math.ceil(w/20/4)*4. why we do like that. could you help to explain thank you!

Yijunmaverick commented 6 years ago

@suke27 Sorry for the late reply. It is not a must and you can just input your original image. Sometimes I found that style patterns (e.g., if the style image has a frame) keep appearing at the image boundary area which looks not visually pleasing. So I pad a few pixels and finally remove it after transferring.

hello-trouble commented 5 years ago

hello, I am very insterested in the work of MultiTextureSynthesis. Is there pytorch code of this paper ? Thank you very much !