search

LucasSheng / avatar-net

Avatar-Net: Multi-scale Zero-shot Style Transfer by Feature Decoration
https://lucassheng.github.io/avatar-net/
178 stars 38 forks source link

question about your paper #4

Closed dypark86 closed 6 years ago

dypark86 commented 6 years ago

i have some question about avatar-net.

  1. training of decoder.

    • in this paper, decoder train only MS coco data set for reconstruct without feature transfer module
    • why did you training MS coco data set only?
    • for example, adain train MS coco data + wiki art data set with adain module.
    • there are any reason for train MS coco reconstruct without transfer module?
    • is transfer module too slow for train ?
    • transfer module can't back propagate?

  2. future direction

    • in future direction, replace style decorator by learnable modules for increased alignment.
    • what is the effect of increased alignment for output image?