Image Completion with Deep Learning in TensorFlow | Hacker News Reader

Image Completion with Deep Learning in TensorFlow

bamos.github.io

229 pointssemanser10 comments

Oh man, those facial images are deep in the uncanny valley. Maybe on the rising side of the slope now, but still way down there.

The photos at the bottom look like posters from an 80s horror movie! The photos with vacant eyes are hella creepy...

Have you seen the animation at the end of the article, where progressively better completions are selected? Most of them seemed pretty natural to me.

Yeah, I saw them - a little difficult to tell with the loop, but they still seem 'off'. Buh.

Thanks for the feedback, I just updated the animation to pause at the beginning and end to make the completions clearer: https://github.com/bamos/bamos.github.io/commit/f6a152851355...

I was impressed with the Labeled Faces in the Wild (LFW) facial auto-completion results, especially since the system was not trained on LFW at all! The results seemed almost too good to be true. Perhaps this is a testament that there isn't that much diversity in human faces?

Very well written overall, and I appreciated the author's thoughts on TensorFlow+torch at the end of the article.

Adversarial training is a fascinating idea, and I love the sound of it. I'd like to start applying that concept in the future.

On a similar vein...

http://www.pyimagesearch.com/2016/08/10/imagenet-classificat...

Uses TF deep learning to classify an image.

It would be cool if you could use a GAN to generate images merely by providing an object name. You could train the GAN on images obtained by searching for that object name in an image search engine.

Awesome! Could use this technique to create a personal version of Google photos (thinking specifically of their automatic generation of albums of 'sunsets' or 'dogs').

Could this be used to generate unique images from a training set?

Reply on news.ycombinator.com