quality on stylegan_ffhq

[IQ17]

Hi, thanks for the paper and the results are impressive!

I tested the code with "stylegan_ffhq" model and "stylegan_ffhq_pose_boundary.npy or stylegan_ffhq_pose_w_boundary.npy", with the default settings, but the results are not very good.

The person identity, age, even gender changed simultaneously with the pose. Regarding to the "stylegan_ffhq_pose_w_boundary.npy", the degree of pose changes are more or less ignorable.

python edit.py -m stylegan_ffhq -o results/stylegan_ffhq_pose_w_boundary -b ./boundaries/stylegan_ffhq_pose_w_boundary.npy -n 10

Is there anything that I have to adjust?

[ShenYujun]

How about setting --latent_space_type=w, since you are using the boundary from W space.

[IQ17]

Yes, I tried with --latent_space_type=w or --latent_space_type=W but the outputs are not face anymore.

[IQ17]

I use stylegan_ffhq to generate the images, and the generated images are great. Thus I think the model is correct.

[ShenYujun]

That is because the current codebase will randomly sample w codes subject to Gaussian distribution instead of the actual distribution. There are basically two solutions (1) Run generate_data.py first, which will save w.npy in the output folder. Then use -i option in edit.py to load the generated w codes. (2) See HiGAN which provides a more robust generator, which will sample codes from the actual distribution.

[IQ17]

Thanks! Using the solution 1, the results are good!

[xing-shuai]