What's the difference between the DIM and RWP ?
Opened this issue · 4 comments
dreamer121121 commented
jsjxyjl commented
hi ,Do you have any other problems with your training?
yucornetto commented
Hi, if you are training the model for DIM benchmark, please use DIM config, if for real-world portrait images, please use RWP config and the data preparation.
dreamer121121 commented
why comp weight set to 0 in config file, why not use comp loss to train network
kartynnik commented
why comp weight set to 0 in config file, why not use comp loss to train network
As mentioned in the paper, during RWP training the composited input is heavily augmented (e.g. with noise), whereas in the compositional loss it would be compared against a blend of clean, non-augmented source background and foreground, which makes no sense.