Hacker News new | past | comments | ask | show | jobs | submit login
This Waifu Does Not Exist (thiswaifudoesnotexist.net)
175 points by hardmaru on Feb 20, 2019 | hide | past | favorite | 23 comments

How much porn was in the dataset? The generated faces seem to have a tendency to have white goo arond the mouth...

~9%. (I didn't think it would be much of a problem since they'd be cropped down and anyway, all my previous GANs sucked too much to care about fine details.)

Was this pun intended?)

It is an occupational hazard.

Note that if you don't like that interface, there's another, for those made of sterner stuff: https://www.obormot.net/demos/these-waifus-do-not-exist.html

Another cool demonstration: training on a particular character tag, so you can e.g. sample from the space of possible Asukas. https://twitter.com/gwern/status/1094728538396205057

I imagine that tools such as this will be invaluable for game developers of the future.

I'm new to this how "This X Does Not Exist" thing ("This Person Does Not Exist" in seems especially impressive). It seems like this images are generated through some Machine Learning Magic(TM), but could anyone here give a rundown of how it all actually works? These all seem freaky good for something generated by a computer.

They are generated by Generative Adversarial Networks.

Have you heard about convolutional networks? They are neural networks that are especially well-suited for problems that have spatial structure (such as 2D images) and translational invariance (a face is a face, no matter of its coordinates in the picture).

Generative Adversarial Networks, or GANs for short is an architecture that arranges an image-generating network and fake-detecting network in a competitive situation. That makes them co-evolve towards generating pictures that are harder and harder to distinguish from the real ones.

Wow, that's really interesting! I've heard a bit about convolutional networks, but I've never heard of pitting two systems against each other like that. Thanks for taking the time to write that out.

For considerably extra lulz, I have added GPT-2 text snippets with an anime plot prime.

They are absolutely hilarious: http://0x0.st/zoT2.png

Some of these generated images look oddly male, or at least gender-unspecific, with their longer faces, eye-nose distances and wider chins. Mixing lots of female drawings inside a gann apparently doesn't result in females in every case.

It has some hiccups beyond gender. For example: https://imgur.com/a/d3SfrnQ

It looks like a bug alright, but then again... it's not something I wouldn't expect to see in actual anime :).

I didn't filter out male faces (no convenient way, anyway), so there are males in it. Not a ton, certainly, but at least a few thousand.

A lot of anime waifus have androgynous features (often intentionally, e.g. Felix from Re:Zero). I wouldn't be surprised if the training set included many such examples.

This is hilarious and awesome.

The size of the generated images seems to vary. Any idea what might be the cause of this?

I am about halfway through upscaling them from 512px to 1024px with waifu2x.

Neat twist on the other website.

Applications are open for YC Winter 2023

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact