Hacker News Reader

Comment by andreyk | Hacker News Reader

I wonder how long the novelty of DALL-E will persist. HN seems to upvote anything titled "I did X with DALL-E". This is a fun post, but it's not that interesting or surprising. Still worth a look don't get me wrong, but personally didn't learn anything new from it. (eg recreating the famous pink Floyd cover with "Outline of prism on a black background in the middle of scene splits a beam of light coming from the left side into rainbow on the right side" unsurprisingly worked somewhat well).

View on HN

"DALL-E" in titles is about to be replaced with "Stable Diffusion." The beta website is already live but the interesting part will be specialized fine-tuned models based on public weights. There should be more technical experimentation since Stable Diffusion weights are only a few GBs and inference can be run on any recent GPU. There might also be more controversies because it can create more uncensored images.

Somebody posted a nice comparison between DALL-E 2, Mid Journey, and Stable Diffusion: https://twitter.com/fabianstelzer/status/1561019187451011074.

I dunno, I don't see a lot of examples on that page where Stable Diffusion outperformed DALL-E 2.

I'm not saying that it does, just that like you definitely hear more about Blender than about Maya on Hacker News, you will hear more about an open source model than a commercial one.

I actually thought there were some interesting things revealed here. Particularly interesting to me were the Nevermind and Abbey Road results.

Nevermind because it showed this weakness in the model in understanding what a dollar bill is. The most novel result being the image where the baby's visage appears inside the dollar.

Regarding Abbey Road, I found it interesting that the model's concept of a public person spans their lifetime evidenced by the images where the contemporary images are used. Also interesting to me is the model's weakness in understanding specific people.

Then again though, I haven't been clicking on every DALL-E post so maybe this is old news.

For me, what was interesting to see was that it didn't manage very well with the Abbey Road question.

I had imagined that some parts of the training data would consist of the actual image, and it would find a good match for it somewhere deep into it's artificial consciousness.

It has no concept of The Beatles or Abbey Road or zebra crossings or perspective or creative composition in photography, so effectively it just mashes up some results from an image search.

If it created something very close to the original it would be overfitting. So it flails around rather randomly in photos-of-Beatles+Zebra+Abbey Road space. The results are novel, but they're also monstrous, distorted, and unartistic.

> unartistic

Says who? Maybe you don’t like the style, maybe it is derivative, but it is definitely artistic, as you allude to by saying it is novel. Monstrous and distorted can be a style.

“mathematical art, 1924, litography, abstract generative art” generates https://pbs.twimg.com/media/FanUkREXEAEH6su.jpg — Can you pick the influences? I can’t.

"low poly game asset, Cthulhu monster, 2000 video game, isometric view" generates https://pbs.twimg.com/media/FanTf3IXgAETece.jpg

I thought https://pbs.twimg.com/media/FanNq6sXkAENyU5.jpg was really interesting, because the LEGO logo is not a 100% faithful reproduction even though the LEGO logo is so ubiquitous - showing that the images are generated.

From a longer tweet thread by @fabianstelzer comparing DALL-E 2 vs Midjourney vs StableDiffusion: https://threadreaderapp.com/thread/1561019187451011074.html

Maybe it would have done better if the prompt had been "Four men - George Harrison, Paul McCartney, Ringo Starr and John Lennon - striding across a zebra crossing situated on Abbey Road, outside EMI studios in London."

It will probably get boring the same time HNers stop confusing “hmm this isn’t that interesting guys” as notable commentary that needs to be shared.

DALL-E generate us some pics of unhappy HN curmudgeons

Here ya go:

https://imgur.com/a/kgQtUMf

I don’t know why it’s got the over-18 warning. It’s 100% SFW.

The top left one is clearly an ape in a hoodie, right on point.

Well then it will never get boring, will it... Fair enough, it's just been striking to see how many of these sort of posts have been popping up. Like I said, it's still fun and worth a look.

Perhaps as a blogger I am extra salty when relatively low effort stuff gets upvoted over things that take a lot of work to write (in some cases, those being things I have written). But hey, that's life.

Low effort stuff can (and often is) better and/or more entertaining than high effort stuff.

Can't argue with that!

To me the most interesting thing about the article was actually just how bad the results are. It's interesting that you picked the Dark side of the moon one, as to me that seemed by far the worse one - none of the pictures really resembled the simple geometry of the original. While I understand why recreating the Nevermind cover was difficult, it was surprising that recreating such a simple geometric pattern failed so spectacularly.

The only cover that really worked from my point of view was the Velvet Underground one, and perhaps the Rolling Stones one. Abbey Road came closer than what I thought it would, but was pretty bad ultimately, and the other three really had nothing usable.

Just like images after text, pretty sure once the novelty of images ends we'll move to animations, music, movies, computer games, and so on.

For the past few years, I’ve long considered Hollywood failed adaptations to be a non-issue because so long as industrial civilization exists, media companies are going to squeeze what they can out of IPs. So while Game of Thrones might have ended quite poorly, I figure we’re only a few decades away from a different rights-holder giving it another go.

Thus, I look forward to the AI-generated versions of famous works that deepfake the original cast into speaking (hopefully) better-written dialogue. Imagine when this technology is widespread, fanfic authors rendering their interpretation of works with the descendants of DALL-E. Everyone gets the dream adaptations and sequels and finales they want.

The trouble with this, for me, is that wouldn’t such a world necessarily mean the end of new actors (not to mention new worlds and stories)? After all, why hire unknowns when you can stack the cast with the A-listers of all time?

And if everyone stacks the cast thusly, there are no opportunities for new actors to work with old and be mentored + opportunity to simply works since every bit of work is what adds up to make a journeyperson.

Still gotta populate the training data

I get that, but I find it hard to see why one would look for NEW training data from actors who will never have the opportunity to become as seasoned as their forebears, when you have 100+ years of filmed historical data (among other formats) from which to pull your players?

Unless copyright changes a lot, the A-listers are going to expect a lot of money for a film with their name on the poster and their carefully curated likeness throughout, whether they have to turn up to shoot it or not.

Totally, but their estates won’t be as precious (as we’re seeing with holograms already). And once a certain critical mass is built up, it’s built up — and it’s be foolish not to use it.

Generative art and by extension DALL-E has a very fast attention decay imo. We can't avoid noticing patterns and getting bored by them. This makes art and music fun because the things that stick and stay interesting JUST DO for some unknown reason.

I love this kind of bad humor of HN.

Then you’re gonna loooove these computer-generated “jokes”…

http://joking.abdn.ac.uk/jokebook.shtml

Reply on news.ycombinator.com