Why GeometryEncoder instead GlobalGenerator

[Michel-liu]

Hello, nice job! I have a question, why don't use the encoder(including downsample and resnet-blocks) part of GlobalGenerator with ngf=32 as Geometry Encoder directly, but train a GeometryEncoder with ngf=64. Thanks.

[ty625911724]

Dear Michel:
Since sketches and images are encoded into a share space, there are two GeometryEncoder for sketch and image individually. GlobalGenerator just translates images to sketches.
Thanks!

[Michel-liu]

Thanks for your answer! Is the 32.2k images list used in FFHQ public available? And I wonder about your strategy for selecting training image pairs in the disentanglement process, for example, gender, age, facial covering, etc. Thanks.

[ty625911724]

Thanks for your interests. We will release the training list in the future.