Same idea with StyleGAN-Encoder repo? about idinvert HOT 2 CLOSED

1. There are two major differences, i.e. domain-guided encoder and domain-regularized optimization. (a) stylegan-encoder does not actually have an encoder to map a given image back to the latent space. By contrast, our encoder provides a much better initialization. (b) We also involve the encoder in the optimization process as a regularizer, to ensure the code to be in-domain. BTW, Image2StyleGAN, instead of this repo, is much like stylegan-encoder (They use exactly the same process, and the only difference is the loss function). As for your mentioned "conducting the same thing", I think all GAN inversion approaches do the same thing, which is to recover the input image with latent code(s). Differently, we raise the problem of exploring the semantic property of the inverted codes. This distinguishes our work from existing work.

2. We explain this in the paper. You are right that it is the domain-guided encoder to ensure the semantic property. The intuitive explanation is the training manner of the encoder. During the encoder training, we will feed the output of the encoder back into the generator to synthesize a photo-realistic image. In this way, if the code does not land inside the original latent space, the synthesis quality will suffer. Alternatively, the code should align with the semantic knowledge learned by the generator for image synthesis. That is what we call in-domain. Fig.3 in the paper also demonstrates this point. To maintain the code to be semantically meaningful, we also involve the encoder in the instance-level optimization. The demo video shows the superiority of our approach over the optimization-based method (like stylegan-encoder and Image2StyleGAN). With the help of the domain-guided encoder, the inverted code is more manipulable.

from idinvert.

BTW, we have another work I think you may have interests in. That work proposes a different optimization-based pipeline by composing multiple latent codes for a single image.

from idinvert.