Hypnogram: Create an image from text | Hacker News Reader

Hypnogram: Create an image from text

hypnogram.xyz

109 pointsandrew-ld48 comments

I highly recommend checking out diffusion networks for image generation. It's hard to put in words the difference between diffusion and something like VQGAN, but the images have a much more organic quality. Here's a mish-mash of images [1] I generated recently. A few (like the bears) used initial images to get the process started, but most are from just a text prompt (the large image of the dog at the bottom right was the result of using another network to upscale one of the outputs).

The results need to be cherry picked, but I'm frequently surprised by the good images. For instance, after some attempts I got this nightmarish result for "my neighbor totoro" [2].

edit: It can also be very satisfying to see the process unfold throughout the iterations, starting from noise and eventually settling on a very convincing imaging [3].

https://ibb.co/6mNTMjH [1] N.B: Probably SFW, but some images are oddly disturbing

https://ibb.co/7KCvXPG [2]

https://imgur.com/a/HEBM0uq [3]

Pleasantly surprised with what I got for "misty mountain":

https://i.postimg.cc/ryDWKPXT/image.png

I wonder, what stops someone just using one of the images generated here as their own "art"?

Nothing. Past cases have settled that the copyright goes to the person who used their creative process to sort through the computer generated images and pick the quality out.

So by typing misty mountain and selecting it as good, you are most likely the copyright owner of that image.

Oh interesting. Do you have a source by any chance?

I did some more searching and it looks like its actually a little up in the air right now but that proposals are looking at making it work this way. So the AI would be considered something like photoshop where if you use the tool to create something, you are the copyright owner.

That the result may be considered a derivative work of something in the training set.

And to give my hot take: My current thinking is that the law should take a strict stance here, as the produced images will compete with, and may substitute for the original ones.

What are the original ones? There is likely no previous image that looks anything like this generation. When you draw a picture, you aren't required to credit every piece of art you have seen in the past which inspired your style.

Someone entered "unreal engine hyper-realistic retro futuristic cityscape" and the results were a fantastic match:

https://i.ibb.co/BVSFHpN/image.png

My own query, "clown factory" was a bit less obvious:

https://i.ibb.co/TYDmRT1/image.png

I think the ability of this model to create truly horrifying pictures is remarkable.

While waiting for my query to generate, I was graced with "Hacker News", felt it was worth sharing back to this thread :)

https://imgur.com/a/OKW411E

I got something more ominous for "hacker news": https://i.ibb.co/hYVPh6k/hacker-news.jpg

I don't want to be overly pessimistic, but I have a strange feeling about AI image or text generators. Who can prove that it actually does what it says?

I have a conspiracy theory: Somebody found a way to quickly generate pictures in this style in a semi-manual fashion, and hired a bunch of people to fake this, mechanical turk style. I have never found an interactive generator that completely convinced me, neither for images nor for text. Either the results are really bad, or the responses take so long that a person could have made it.

But why would somebody fake something like this? There is a lot of money in AI. It seems totally plausible that someone invests five or six figures to set up a fake demonstration, in order to raise millions in investments.

Now I want to make clear while I get a big Theranos vibe from these kind of demos, I don't want to accuse anybody concretely. I think a lot is technically possible and I do believe a lot of research is real, but I also believe some demos are "fake it till you make it" territory.

I might be getting whooshed here, but if you're serious, there's a couple colab notebooks floating around, you can look at the code that generates it or run it on your own hardware. Here's one: https://is.gd/artmachine

Thanks! I'm only half joking. I used traditional machine learning in my PhD but never got into modern deep learning techniques. I tried a couple of collab notebooks and tutorials but never succeded with anything more complex than NIST numbers. I think over the holidays I'll give it another shot because I find this stuff really interesting (art generation, style transfer, but also transfer learning in general).

The conspiracy theory is more a jab at the fact that we live in a world with Theranos, NFTs, and clones of Dogecoin. There actually have been a couple of high profile AI fraud schemes, like Augustus Intelligence, where people collected investments and government money and probably never had a product.

You can download the code and run it on your computer, without internet. Or maybe you can’t, and I’m just part of the conspiracy. Only one way to find out!

David Weinberger has written here: https://aeon.co/essays/our-world-is-a-black-box-predictable-... about this difficulty of thinking about such systems - as others say, you can look at the code, but I understand the resistance to just accepting it.

it's genuinely interesting to see the weird mix of random nouns, nerd things (not surprising considering what site i'm posting this on) and people trying very hard to get AI porn.

I put "all star by smash mouth in the style of the last supper" and was not disappointed (https://i.imgur.com/3flZ0CN.jpg)

Nicely executed. Showing other people's results while you wait is great. What's your throughput? (images/sec)

Finally got to make it-- just had to trim the words some.

https://news.ycombinator.com/item?id=29004677

https://ibb.co/rvtFDR9

The would be great to illustrate a markov-chain generated blog.

I put "Starship landing on the moon" and "Starship landing on Mars". Let's just say that the results generated were far better than I would have produced given pretty much any medium to create in.

I won't say what I put in b/c I don't want my IP associated with my HN account, but it's basic and it keeps failing.

I did get to see something about "Jew money" scroll by along the bottom while I was waiting, so, thanks internet.

You put in your IP address?

No, if I say what I "searched" for they'll be able to relate the IP address that requested it to this account.

i know that feel, bro, i am paranoid as well

Someone really needs to put a filter on that "latest images" section..

That said, this is impressive and extremely addictive!

For those interested, I'm posting some of my favorite results on IG (@gantagonist)

Some of these are very cursed.

Yes.. "absolute nightmare terror photorealistic disney unreal lovecraft style"

https://ibb.co/Wyhcq3T

Imagine using a future version of this to generate a VR game.

Trained on photos from your childhood