back
100 comments
This is an amazingly beautiful hack. I'm kind of surprised we never saw something like this done by hand before the advent of SD -- plenty of QR code art, but none of it like this. I'm impressed.
There are quite a few doing them by hand, eg https://dribbble.com/shots/3658816-Art-QR-Code-Design

I seem to remember some NY artist trying to sort of vaguely claim/intimate invention (falsely) by having an exhibition too. Ah yes, here https://art-qrcode.com/

It's costly to make such images, that's about it. Other way around you'd have to decide on the QR, wait for the commission, pay a substantial amount of money. Probably wouldn't even get such subtle codes either, this seems like straight up CV-trickery.
It's for sure brilliant.
Something I learned recently: the phrasing "QR Code" is a trademark of a company called Denso Wave, who invented them. While the method of producing and reading QR codes is free for all, (Denso Wave waived their rights to the patents) they still protect their trademark. This has lead many Japanese companies that use QR codes, to call them publicly something else, such as "2D codes", in fear of getting sued.
Huh. Turns out it's a subsidiary of the same Denso that's a massive automotive supplier and largely owned by Toyota. Asian conglomerates are so interesting.
IIRC they created them in the first place to make manufacturing easier
Nothing compared to european conglomerates, which do wastemanagment, auction horses and build fighter jets...
I have seen them called "Quick Scan Code" or something similar in some ads.
Interesting bit of info, thanks for sharing. Sadly for them, I think the genie is out of the bottle on this one; while I've seen a few people use alternate names, most call them QR codes.
I love how every news outlet reporting this will essentially be forced to link to the author's site if they want to include the image.
Imagine if some construction company created a whole neighborhood that from the right vantage point would be a valid QR code.
It's all fun and games... Until someone forgets to pay the annual domain fee.
Fun to imagine in the abstract. I used to take classes in a building designed to have a certain appearance from above, and the user experience was definitely the worse for it

    1. Secretly repaint one wall of a house and redirect the qr code to a slightly different url
    2. ???
    3. Profit
Old Maxtor campus in Milpitas was supposed to look like an HDD https://goo.gl/maps/JbRWqZseAVL14n167
If you zoom out they are very recognizably QR codes. A commenter made the point that what you're seeing in that case is the low frequency component of the image. Kind of like the Marilyn Monroe / Albert Einstein illusion.
This kind of QR codes should work nice on big billboards
Beautiful work!

From the title I was expecting the QR codes to move, but they're actually "anime drawing style", not "anime"

Does that mean the accurate term would be "manga"?

I've never been sure if the only difference between the two is still frames vs motion, or something else.

> From the title I was expecting the QR codes to move, but they're actually "anime drawing style", not "anime"

Given the description of the apparent technique, its likely trivial to extend it to animated QR codes using the same ControlNet models and ControlNet inputs and existing techniques of animating travel through prompt and/or seed space for StableDiffusion.

the scanned result redirects to qrbtf.com, which OP ported some QR libs into javascript [1]

notably they also implemented the "Halftone"[2] style.

[1] https://github.com/ciaochaos/qrbtf

[2] http://vecg.cs.ucl.ac.uk/Projects/SmartGeometry/halftone_QR/...

That is impressive. I know the images will have been selected for the fact that they are particularly good examples, but I've seen a number of implementations of trying to mix a photo into a “standard blocky code” (the girl and Trump examples on that page) and none previously produced results quite that good. I shall have to have a play.
> Next, they fed existing QR codes into the Stable Diffusion AI image generator and used ControlNet to maintain the QR code's data positioning despite synthesizing an image around it, likely using a written prompt.

A few months ago, after seeing that ChatGPT could produce reasonably accurate MD5 hashes, I tried generating QR codes with a Midjourney prompt. It failed in a really interesting way - it generated _symmetric_ QR codes, even though QR codes are explicitly designed to be asymmetric (think of the timing pattern, the three large blocks). So while I could claim ChatGPT "understood" MD5 hashes, I can't claim that Midjourney "understands" QR codes.

How long until generative models can produce working QR codes on their own?

Forget QR codes, until very recently Stable Diffusion and Co could not generate legible text in images.
Now, who wants to combine this with the tutorial from https://news.ycombinator.com/item?id=36173441 and decode one of those by hand? :)
This would pair excellently with a WiFi QR code for restaurants showing their Wifi password. (eg: <https://qifi.org>)
For the love of sanity, please don’t. In real life, I struggle enough to scan normal QR codes when the redundancy of error correction is here to… you know… correct errors and help the scanning process.
>This discovery opens up new possibilities for both digital art and marketing.

Like hiding malicious URLs in public.

Hah, even without the hacking angle, having URLs to promote shit popup when you point your camera somewhere could be the next obnoxious advertising.

For a few years now phone cameras have been able to automatically detect QR codes, maybe their next feature will be an option to turn off "smart features"...

The alignment squares here are a pretty big giveaway, and it seems like it'd be really difficult to fix that without making them unscannable
Or making QR codes harder to read on certain cameras or lighting conditions by removing the error correction!
I wonder if this would work by simply using out-painting to preserve the black dots, to make a background, and then use in-painting on the dots with the condition that they stay dark
Yes, both white and black dots though. So long as the 'central' 30% ish of each grid location is appropriately 'light' or 'dark', it will scan.
The use of controlnet to preserve the important parts of the QR code was genius here.
The internet is flooded with these in one day, but at least this one does not work for me (iPhone 12 mini), it does detect the face of the lady.

Edit: Ok, it does work at a larger distance.

I believe this will be widely adopted and quickly too! Fantastic work!
Working seems like a stretch. I can't get any of them to scan.
Which app are you using? For most of them I needed to get a few feet away, as mentioned in the article. For some it seemed I needed to be more exactly facing it, not angled. I got all of them to scan pretty easily.
Works mostly fine for me from Firefox on Android. All scans fine except one.
Working for with iPhone's scanner.
I guess the next step will be someone who converts themselves to a living QR code via plastic surgery and makeup.
My phone (Samsung S23U) couldn't read the first image.

It found the face, it found the text on the page, it highlighted sections in case I wanted to copy the image/text, but no QR response.

I would absolutely love to see these instead of regular QR codes everywhere
This is so insanely cool I don't even know where to begin.
You could try scanning it