Code for the demo used in blog post.

Setup

Run ./script.sh.

The script install pip packages, downloads pretrained model weights, manipulation vectors and a facial landmarks detector for aligning input faces.

Using pre-trained model

To use pre-trained CelebA-HQ model for encoding/decoding/manipulating images, use model.py

If your image is not aligned, use align_face.py to align image.

To create videos, check videos.py

Create manipulation vectors for an attribute of your choice

Scrape images from the internet for an attribute of your choice (say red-hair vs not red-hair). Then, to obtain manipulation vectors from, use get_manipulators.py

To see how it was done for the CelebA-HQ dataset (which has 40 attributes), first download the input images (x.npy), their attributes (attr.npy) and their encoding (z.npy)

curl https://openaipublic.azureedge.net/glow-demo/celeba-hq/x.npy > x.npy
curl https://openaipublic.azureedge.net/glow-demo/celeba-hq/attr.npy > attr.npy
curl https://openaipublic.azureedge.net/glow-demo/celeba-hq/z.npy > z.npy

Then, run get_manipulators.py

Run server and client for demo

To run server for demo, run python server.py.

To run client, run python -m http.server (starts a local http server at port 8000) and open 0.0.0.0:8000/web in your browser.

To test client, upload test/img.png. You should see aligned image and be able to move sliders.