Style generation...

[rosinality]

Style generation...

[zhibo-liu]

@rosinality
It seems that truncation trick in W space is not used during training:

style-based-gan-pytorch/train.py

Line 161 in b6898ed

fake_image = generator(gen_in1, step=step, alpha=alpha)

here's what I got when running pretrained 256px model without truncation :

here's with truncation (unchanged generate.py):

Apparently, images without truncation are in poor qualities. Why does this happen? Is it because that the pretrained model needs more training? Or we can only generate images using truncation tricks?

Thank you

[rosinality]

It is used for style truncation to reduce deviations of styles from calculated mean styles.

[MichaelMonashev]

Did you try code like this:

normal_distribution = torch.distributions.normal.Normal(0.5, 0.1)
input = normal_distribution.sample(sample_shape=(1,512)).clamp(min=0.3, max=0.7)
style = generator.style(input)[0].to(device))

0.1 , 0.3 and 0.7 - parameters for tuning deviations.

[rosinality]

I haven't, would it better than truncation on the latent codes from mlp?

[MichaelMonashev]

I correct code above. I suggest it after reading this https://github.com/soumith/ganhacks#3-use-a-spherical-z .

[MichaelMonashev]

The basic idea is to sample from a normal distribution very close to 0.5.

[rosinality]

@rentonliu I think longer training will help, as the intuition of the truncation trick is that image quality is poor because of less training in the areas that has low probabilities.

[rsn870]

Can you please clarify what do you mean by areas that has low proabiliites ?

[rosinality]

Some area of noise spaces has relatively lower probability than rest of the spaces. So during training samples from that area will receive less training signals because of less samples came from that area. So quality of samples from that area will be lower than higher probability area. Truncation trick is used for pull samples from these lower probability area to the mean of latent spaces.

[rsn870]

By noise spaces are you referring to the gaussian noise injected during training or the latent distribution ?

[rosinality]

@rsn870 Basically for noise inputs to mlp, but I think latent distributions from mlp also behave similarly.