Navigating StyleGAN2 w latent space using CLIP

an attempt to build sth with the official SG2-ADA Pytorch impl kinda inspired by Generating Images from Prompts using CLIP and StyleGAN based on the og projector.py

things learned:

• it's better to generate initial w values from a well converged sample rather than starting with random or median ones
• optimizing w and noise inputs works better than w alone
• default values of 0.02 for LR/noise work fine with portraits

Quick start

• clone SG2 repo, copy clip dir from CLIP repo, install pytorch 1.7.1 and stuff

• pick a suitable SG2 PKL (eg FFHQ)

• pick a seed

• run python3 approach.py --network network-snapshot-ffhq.pkl --outdir project --num-steps 100 --text 'an image of a girl with a face resembling Paul Krugman' --psi 0.8 --seed 12345

• alternatively, one can start from a w vector stored as .npz python3 approach.py --network network-snapshot-ffhq.pkl --outdir project --num-steps 100 --text 'an image of a girl with a face resembling Paul Krugman' --w w-7660ca0b7e95428cac94c89459b5cebd8a7acbd4.npz

FFHQ test

python3 approach.py --network stylegan2-ffhq-config-f.pkl --outdir ffhq --num-steps 100 --text 'an image of an Instagram influencer girl' --psi 0.7 --seed 32