Nightmare Machine – Horror imagery generated by deep learning | Hacker News Reader

Nightmare Machine – Horror imagery generated by deep learning

nightmare.mit.edu

150 pointsnecrodome31 comments

The initial "deep dreaming" images were far scarier. Nothing compares to animal faces literally popping out of the woodwork, as if on LSD.

Automated generation of synthetic, photorealistic human faces was mildly terrifying as well.

Still, this is a cute Halloween project.

In my opinion, some of the images I've seen NNs produce without trying to be scary, were scarier than this. There is something really unsettling about things that look almost real, but aren't quite right.

https://raw.githubusercontent.com/torch/torch.github.io/mast...

Yeah, those are straight up uncanny valley stuff.

I find them rather comforting, oddly enough. It's the humanity that terrifies me.

Agreed with others, they look like they're supposed to be scary, but something is missing?

A few days back I see (yet another) article about Stephen Hawking's concerns for humanity and AI's potential for harm. Today I see we're teaching AI how to scare us.

Super :)

I know it can be computationally expensive, but larger images would really help in this case.

Hi there! We actually have higher-quality versions of these images but didn't want to overwhelm the main page (and users' bandwidth). But we can put the larger images to the sub-pages, thanks for the suggestion!

Thanks for the response.

What about putting up higher res versions when you click on the "thumbnails"?

Like here: http://nightmare.mit.edu/#portfolioModal30 (can't link directly to the modal)

The computed stylized can be in higher res and only query the server once the user clicks on the thumbnail. maybe even the original in higher res. I mean 1920x1080+ and maybe some options to downsample to specific res.

Another suggestion is to figure out a way to maximize artifacts. My scariest experiences with deep learning have been with unnatural mutations of natural objects.

Something like https://zo7.github.io/img/2016-09-25-generating-faces/random...

Or some results from this: https://github.com/Newmu/dcgan_code

I'm sure one can take it further.

Deep torture, anyone? Gradient descent on stimuli to get what you want out of adversaries. Kind of scary

On behalf of Saito Game Team, I most humbly apologize.

For anyone who doesn't get this, watch Black Mirror.

rot13: V jngpurq gur rcvfbqr, bcrarq UA naq gurer vg jnf. Ba pbaprcghny yriry vg jnf bar gb bar znccvat orgjrra gur rcvfbqr naq vzcyrzragngvba.

The more time goes on, the more confident I become in the notion that I am currently living in an episode of Black Mirror.

Anyone else think that the haunted faces bear an uncanny resemblance to the portraits of Francis Bacon? Which, unsurprisingly, are quite nightmarish in many respects:

https://www.google.com/search?q=francis+bacon+portraits

It would be a lot more terrifying if it somehow tried to merge the photos skulls and other traditional horror imagery (the taj mahal photo refection could easily be a skull).

Great idea! We actually tried the skull style on some pictures before (https://www.instagram.com/p/BLcU-MzBoEe/?taken-by=nightmare_...) and they seem okay. We will try it on buildings and see how it turns out!

My I have a request? My friend, after seeing the images, she asked if it is possible to make Borobudur becomes a scary place.

Here is the link: https://en.m.wikipedia.org/wiki/Borobudur

Thanks

Check out deepdreamgenerator.com -- requires a free account, and there are some usage limits, but it's a lot of fun to play with, and can do similar NN-based styling of images.

This reminds me of an episode in the new season of Black Mirror when they connect a guy to a device which data mines his memories and conjures up horrors to entertain the guy...SPOILER ALERT!!...and he dies 0.04s after the connection.

For me, the grid of faces is much scarier than each on individually...interesting.

You might want to check out the site of Chet Zar for more scary faces (artwork): http://www.chetzar.com

I wonder where the bulk of their training images are from? The website doesn't seem to indicate it. These remind me of the parallel universe in "Stranger Things."

This is disturbing :) Elon is right, AI is evil!

I wonder what artchitecture was used, inception, RNN or a GAN?

It is perhaps ironic that I find the fright-night theme beautiful and soothing. Would love capturing them for real on camera.

we may end up generating our own cthulhu

Why? Is there any remote need to generate scary faces? Or is it done just to have catchy PR title. I hope there is some real motivation to do this line of research.

Fun?

Is MIT for fun?

Ideally, at least half for fun, yes.

Reply on news.ycombinator.com