Question regarding the encoders

[jis478]

Hi, thanks for the awesome paper. I have some questions regarding the networks in the paper.

Is there only a single generator covering both domain A and B? How is it supposed to generate both domains' features? (For discriminator it has two branches so it makes sense.)

Thanks,

[bing-li-ai]

Many thanks for your interest.

Yes, the weights of the generator are shared between A-->B and B--> A. According to our observation, such single generator can generate high-quality results, if loss functions are properly designed. Many recent works such as FUNIT and stargan also only use single generator for two/multiple domains.