Comparisons between PATN and PATN_Fine

tengteng95 / Pose-Transfer

Code for the paper Progressive Pose Attention for Person Image Generation in CVPR19 (Oral).

MIT License

735 stars 160 forks source link

Comparisons between PATN and PATN_Fine #66

Open Honlan opened 4 years ago

Honlan commented 4 years ago

Hello. Thanks for your awesome work.

I found that the generator of PATN uses instance norm, while that of PATN_Fine uses batch norm (models/networks.py L119 define_G).

What's more, PATN uses lsgan while PATN_Fine doesn't (--no_lsgan).

Could you explain for me why you make such changes?

Thanks a lot.

jessemelpolio commented 4 years ago

Our initial intention is trying to stabilize the training process and reduce artifacts. These techniques are of some help to these aspects but show no further potential to improve the image quality quite a lot since batch norm is not sensitive as instance norm in capturing the style statistics of a single person.