search

Prinsphield / ELEGANT

ELEGANT: Exchanging Latent Encodings with GAN for Transferring Multiple Face Attributes
https://arxiv.org/abs/1803.10562
MIT License
263 stars 42 forks source link

Can ELEGANT perform inter-class translation? #19

Open imlixinyang opened 4 years ago

imlixinyang commented 4 years ago

Another question is that it is shown that ELEGANT can perform multi-label translation, but I wonder whether it is performed by two-step translation or a single step (exchange two codes in or not in the meantime)?

Prinsphield commented 4 years ago

I guess the question that you are asking is: can we swap the hairstyle of two images with bangs? I did not try that, but I think ELEGANT can do that.

It is done in a single step translation. We claim that ELEGANT can control multiple attributes simultaneously. See Fig. 9.

imlixinyang commented 4 years ago

Thank you for your reply, I'll have a try later. It would be amazing if it works because these two tasks are not involved in the training.

Jessicall commented 4 years ago

I guess the question that you are asking is: can we swap the hairstyle of two images with bangs? I did not try that, but I think ELEGANT can do that.

It is done in a single step translation. We claim that ELEGANT can control multiple attributes simultaneously. See Fig. 9.

hello @Prinsphield About the second question, it means exchanging two attribute codes of two reference images with the code of input image in the meantime? I wonder how is this achieved? because the model is performed on a reference image and an input image.

Prinsphield commented 4 years ago

I guess the question that you are asking is: can we swap the hairstyle of two images with bangs? I did not try that, but I think ELEGANT can do that. It is done in a single step translation. We claim that ELEGANT can control multiple attributes simultaneously. See Fig. 9.

hello @Prinsphield About the second question, it means exchanging two attribute codes of two reference images with the code of input image in the meantime? I wonder how is this achieved? because the model is performed on a reference image and an input image.

Hello @Jessicall, since different attribute codes are disentangled represented in the latent encodings, so we can exchange the encodings of the input and two reference images simultaneously. You may refer to the code for details.

Jessicall commented 4 years ago

Thanks for your reply!