search

sunshineatnoon / LinearStyleTransfer

This is the Pytorch implementation of "Learning Linear Transformations for Fast Image and Video Style Transfer" (CVPR 2019).
https://sites.google.com/view/linear-style-transfer-cvpr19/
BSD 2-Clause "Simplified" License
380 stars 87 forks source link

[code] Question about the parameter #16

Closed diaodeyi closed 2 years ago

diaodeyi commented 2 years ago

The style weight and content weight in other style transfer model usually is the former is larger than latter. But in your experiment is inverse, the former is lower than the latter , is there any reason for this, please?

sunshineatnoon commented 2 years ago

Different to conventional practice for the MSE loss, we didn't use average between two gram matrices as shown here. From my experience, this gives better stylization results. This is why the style loss weight is smaller.

As to the specific value of the style weight, it is empirically decided.