Unified GAN for Image-to-Image translation | Hacker News Reader

Unified GAN for Image-to-Image translation

github.com

90 pointsnnx21 comments

Is it possible to download a pre-trained model?

Yes. There are many pre-trained model. You can then use transfer learning to build on a pre-trained model to benefit from the existing computational effort.

Where are the pre-trained models?

In the examples, the source images are very very uniform (Caucasian, young, healthy).

Does this work with other images? Can it age photos of people who are already over 50 for example? Or non-whites? Or children? Etc.

I'd like to see it applied to completely different things, what does it do to animals, a grid, a brick wall, a car, etc.

Is a trained model available anywhere?

I wish researchers showed a random mix of results, not just their best results ...

Some do - they show failure modes.

I frankly don't see why it couldn't add more years to 50y. Moreover, it could likely "unage" over 50y to 20y, just by inverse of their aging method. GANs are unbelievably powerful.

I think this (or something like this) can be used as a way to transmit a video chat on a very low bandwidth (think tens of bytes/second) level.

You're talking about an animated avatar that just happens to be of an actual person? And only in 2D?

For a start, yes. A proof of pricinple would be nice. Would probably look very shitty, but would use very little bandwidth.

I read about a fictional system like this in "A fire upon the deep" by Vernor Vinge a long time ago but never thought it would be feasible. There a scene of an entire room of people interacting was being deconstructed/reconstructed for a video chat using low bandwidth.

Cool as hell. Is this flexible enough that it could realistically be applied to generic selfies and not just these very uniform photos?

Make this into an app and you could have the next Prisma!

You mean FaceApp?

This is fantastic! Great work! Thanks for sharing!

Well, this is horrifying. Thank you. Now everyone that has one photo of me to match against now has several of them.

Well, no, you can't create information that doesn't exist. They don't have any better source to match against your blonde-hair picture than their imagination. The novelty here is that they may outsource that imagination to the machine now.

Is this useful for simulated celebrity and fanfiction porn yet? Would be very interested in that PoC.

The article doesn't point this out, but in GANs, the parameters actually move on a continuous range. In other words, they don't just have several of your pictures, they've got all sorts of movies :

A movie where your face starts out "neutral" and becomes angry. A movie where you start out angry, go to surprised and end happy. And so on.

Excellent progress, I must say.

> they've got all sorts of movies

Here it is: one glorious hour of walk through the latent space of faces. You can "feel" the geometry of the latent space and its meanderings. https://www.youtube.com/watch?v=36lE9tV9vm0

Generated with Progressive GAN in high resolution. The quality is mind blowing, highest ever attained in any GAN. Author: NVIDIA

This is so awesome, thank you. Can't wait to play around with it

racial bias in computer vision strikes again!

Reply on news.ycombinator.com