search

mchong6 / SOAT

Official PyTorch repo for StyleGAN of All Trades: Image Manipulation with Only Pretrained StyleGAN.
MIT License
380 stars 56 forks source link

use e4e latent #3

Open loboere opened 3 years ago

loboere commented 3 years ago

is possible load a e4e latent?

mchong6 commented 3 years ago

If I am not wrong e4e is pretrained on 1024 resolution GAN, so it wouldn't work directly since I am using a 256 resolution GAN. However, my method should generalize to 1024 resolution GAN, so I can see you working with e4e that way.

amadeuzou commented 3 years ago

I used origin latent projected by face.pt with a real world image, and got a nice result:
real world image | projected image as input | generated reference image | output styled image output (2)

mchong6 commented 3 years ago

This is very cool and works better than I imagine. Did you use the projector.py code to do GAN inversion?

amadeuzou commented 3 years ago

yes, the projector.py code in rosinality's project. but i got bad results with your train.py code on disney dataset: 010000 [sample_disney/010000.png]

mchong6 commented 3 years ago

Did you start training with face.pt?

amadeuzou commented 3 years ago

yes, the early model results are normal, such as 1000iters.

mchong6 commented 3 years ago

try using early model then. IIRC, this model I uploaded is only 2k iterations.