Zardinality / WGAN-tensorflow

a tensorflow implementation of WGAN
578 stars 200 forks source link

GAN code to convert synthetic images to real.. i.e. working with paired images #12

Open rohitsaluja22 opened 6 years ago

rohitsaluja22 commented 6 years ago

Hi, please share a WGAN code to convert synthetic images to real (with paired images) if you know. Any other vanila GAN would also be helpful.

Zardinality commented 6 years ago

@rohitsaluja22 Have you learned about pix2pix and its follow-up work cycleGAN? You can find the code in their respected project page.

rohitsaluja22 commented 6 years ago

@Zardinality, I want a simple model with good loss function. In the end, I want samples from real distribution only. These models which you shared I guess give a style transferred image of synthetic version since they try to reconstruct the synthetic image back.

Hope you understand what I mean. I want to try out extreme changes to the synthetic image and am not interested in reconstructing it back.. e.g. would be changing a word image in one font to the same word in another font..

Zardinality commented 6 years ago

@rohitsaluja22 I am still not sure what you mean exactly, but if you want a one-sided version of Cycle-GAN, here is a repo.

rohitsaluja22 commented 6 years ago

@Zardinality check this: https://arxiv.org/pdf/1801.07156.pdf I want to check out if vanila GAN can do this..