Open roimulia2 opened 3 years ago
This is indeed a limitation in the current model. One solution to support multi-resolution inputs is to prepend a global (average) pooling layer before the fully connected layer when transforming the encoder features to the style vector. A workaround without retraining is to run inference on a crop of the full image for each masked region.
Hey, thanks for replying!
Should we expect worse/better/same results doing so or it's unknown?
Hey there, great paper!
I'm wondering what should one do when a higher quality image needs to be the input (full HD or higher). Is it a limitation of the model? Can it be adjusted to support any kind of resolution?
Thank you!