The network output is slightly different from our expectations about pixel2style2pixel

[yuval-alaluf]

Hi @xa0082249956 ,
Can you please clarify what you mean by "the Toonify model provided in the readme as the weight."
Are you training on your 50,000 pairs of (reals, toons) starting from the pretrained model we uploaded to the repo?

[xa0082249956]

Hi, we are training by

python scripts/train.py \
--dataset_type=ffhq_encode \
--exp_dir=experiment_disney \
--workers=8 \
--batch_size=2 \
--test_batch_size=2 \
--test_workers=8 \
--val_interval=2500 \
--save_interval=5000 \
--encoder_type=GradualStyleEncoder \
--start_from_latent_avg \
--lpips_lambda=0.8 \
--l2_lambda=1 \
--id_lambda=1 \
--w_norm_lambda=0.025 \
--stylegan_weights /data/pixel2style2pixel/ffhq_cartoon_blended.pt

And the data are pairs like (real,toon).

[yuval-alaluf]

Cool. So a few points:

1. We train our toonify model using only real images (no paired data) and therefore the loss lambdas we use may not be optimal for your training which you do with paired data. I would consider playing with trying to decrease the w_norm_lambda.
2. In any case, I think the results you got are not bad at all (especially the second one). With that said, playing with the lambda values may give slightly better results since you have paired data to work with.
   Please note that the results you can't may not be as good as the results you get with optimization, which takes a substantially longer time for inference compared to our method.

[xa0082249956]

Great! Thanks for your reply. I will try it.

[brightmart]

Two questions regarding "we use about 50k pairs of pictures generated by Toonify Stylegan to do pairs-training".

1. Where to find the model of 'Toonify Stylegan'?
2. How to do pair-training?

[yuval-alaluf]

Hi @brightmart
The toonify StyleGAN model is provided in the README and can be downloaded here:
https://drive.google.com/file/d/1r3XVCt_WYUKFZFxhNH-xO2dTtF6B5szu/view?usp=sharing

To do paired training, you need to first generated pairs of (real, toon) images. While I don't know exactly how the paired data was generated in this case, you can check out the following Google Colab from Justin Pinkney to see an example of how to generate the paired images:
https://colab.research.google.com/drive/1s2XPNMwf6HDhrJ1FMwlW1jl-eQ2-_tlk?usp=sharing#scrollTo=cuMEHnpmI1Mj