search

yuval-alaluf / SAM

Official Implementation for "Only a Matter of Style: Age Transformation Using a Style-Based Regression Model" (SIGGRAPH 2021) https://arxiv.org/abs/2102.02754
https://yuval-alaluf.github.io/SAM/
MIT License
632 stars 151 forks source link

Time for training #33

Closed jinjiujiujin closed 3 years ago

jinjiujiujin commented 3 years ago

Hello, I am training the network, but my output at 5800 steps is far from satisfied. Here is the output. 87accb8023e21f4a5527293b5d05a4c Could you tell me how long the training will take? My training command is the same as the one in README.

yuval-alaluf commented 3 years ago

Can you clarify what you are not satisfied with? If it's the reconstruction of the background, then in the paper we discuss that preserving the background when encoding into StyleGAN's latent space can be challenging so it makes sense that the background in the top image is not perfect. Are you also using the pre-trained pSp during training for the initial latent embedding?

jinjiujiujin commented 3 years ago

The aging results is not much good. I used pre-trained pSp model for pre-trained encoder initialization and IR SE model for encoder initialization. And my second question is about the training time, I have trained for 7000 steps. How many steps do I need to train next?

yuval-alaluf commented 3 years ago

7,000 steps are not enough for training. You need to train for about 60,000 steps to get good results (that is the number of steps we trained our official model for)