Photo Editing with Generative Adversarial Networks (Part 1)

Originally published at: https://developer.nvidia.com/blog/photo-editing-generative-adversarial-networks-1/

Adversarial training (also called GAN for Generative Adversarial Networks), and the variations that are now being proposed, is the most interesting idea in the last 10 years in ML, in my opinion. – Yann LeCun, 2016 [1]. You heard it from the Deep Learning guru: Generative Adversarial Networks [2] are a very hot topic in…

A nicely written article, thank you for the efforts.

I agree with you!

awesome work! Thanks for writing the auto-encoder to train images to generate the corresponding z values. I am excited to dig into this. Can this can be run through nvidia-docker yet?

The dropbox link used to download the celebA data set has been temporarily disabled due to high traffic. Is there an alternative place to download the data from?

Hello, the main page for the CelebA dataset (http://mmlab.ie.cuhk.edu.hk... mentions an alternative Baidu drive but that also seems to be down, unfortunately. You might want to get in touch with the authors? Sorry about that!

I see you took the initiative of creating an issue on Github, thanks! For the record: https://github.com/gheinric...

I made a basic walk through of how to get a gcp instance running, installing nvidia-docker, and simple setup to run the same example above with nvidia-docker. They are just personal notes, but I figured I would share if anyone was a total newb like me.

Thanks to @enthusiasto:disqus for the Dockerfile!

https://ljstrnadiii.github.io/

Great illustration. It occurred to me that the sample space from which the forger obtains objects is finite. This implies that in all scenario the eventual outcome is that the Expert will beat the forger, will get to a point of perfection (a convergence to 100% accuracy). Is that right?

Greg, would you mind sharing how you created your network architecture pictures?

Hello, I used https://github.com/gwding/d...

Hello, the z-space has a finite cardinality but the possibilities are infinite (ignoring floating-point quantization). Besides, the Expert usually does not have enough capacity (enough neurons) to memorize every single image.

candicegerstner yes