AI human generation gone wrong...

in #blog2 years ago (edited)


This isn't the result of a very low-quality photoshop effort for a scary image. This person you see here doesn't exist. It is basically an effort of AI generating a human that didn't turn out that well... (in this case)

It is an example of Generative Adversarial Networks (GANs), where a computer learns from examples to create something new.

The process is sort of as follows: There are 2 main algorithms. One that will generate new images and another that will compare them to the real examples (that you inputted) and say if they look alike. All the results are then put back into the systems to create better fakes and to also improve the "detective" at identifying fakes. Put this in a long loop and eventually, the images created will pass as real ones.

GANs are pretty fun to experiment with. But given the fact that the technology is built to become good at fooling "the detective", there is a major risk around it. We recently have seen an example where the creators of OpenAI's text generator model have said that they consider not releasing their technology because it became too good and could potentially be used to generate very convincing fake news.

Looking at the picture at the top, this may be hard to believe. That wouldn't fool anyone. But it took me a while to find an example that bad when playing around with Nvidia's StyleGAN. Most of the images were pretty convincing. The 3 images below are also generated and are "people that don't exist". They even managed to create a pretty cute guy that I'd love to date 😉. (So happy that fake people will never feel harassed #metoo)


These people don't exist

I love the possibilities this gives us because it is hugely powerful. They also use similar (more advanced) tech to come up with new medication. It is becoming more and more possible to generate images of very specific things. GAN models exist where you only describe a certain flower: "White flower with pink striped petals" and it generates that image.

With this power comes great responsibility, as they say. It's easy to see that this could put a lot of people out of work, like models, photographers and eventually actors and the big movie stars. With technology becoming cheaper, it will become cheaper to hire a developer once to build a system that generates photos to sell your products, instead of always hiring a model and photographer to do that work for you. I know it would make creating stock photos much easier for me without getting all those model releases. I just don't know if the stock photo website believes me when I say the person doesn't exist...

Anyway... I'm super curious to see where this will go. You too? Give the image generation a spin yourself without building anything: (oh, and DO post the most interesting faces that you got!!)


I do love that Elon's company chose not to release the AI Text generator instead of just rushing to be first to market... but yeah, we might see the world's online population become several degrees larger than the human population. The TV show Catfish might become very, very boring.

haha... could be cool though ;)
But seriously, aren't we already doing that when we introduced Alexa to Google Assistant? ;)

Congratulations @r00sj3! You have completed the following achievement on the Steem blockchain and have been rewarded with new badge(s) :

You got more than 1500 replies. Your next target is to reach 1750 replies.

Click here to view your Board
If you no longer want to receive notifications, reply to this comment with the word STOP

Do not miss the last post from @steemitboard:

Valentine challenge - Love is in the air!

Support SteemitBoard's project! Vote for its witness and get one more award!

it gives me the creeps. I've been using OpenAI's MUSE, to generate text and music and it's unbelievable what they are achieving with machine learning

Posted using Partiko Android