How to insert hidden messages in AI images
Behold the latest AI gimmick
Have you seen the above image on social over the past few days? Did you immediately see the hidden message and think “ah, this is how Boomers are going to get into generative AI”? Or, like me, did you spend an inordinate amount of time squinting at the image, moving your phone further and closer away, like it was a 90’s magic eye illusion. Until finally, when looking at it from the corner of your eye, it became clear that the people’s clothes spell out the word ‘obey’?
Either way, this type of image is suddenly everywhere on generative AI social media. Along with the usual hyperbole that this will change EVERYTHING. This, of course, won’t. But there’s still some fun to be had with it for the moment.
The viral ‘obey’ effort was, unsurprisingly, shared uncredited on The Platform Formerly Known as Twitter (TPFKAT), But it came from Reddit user 3Deal eight days ago in the Stable Diffusion subreddit. There’s a healthy culture on generative AI Reddit of creators showing their workings. A ‘workflow included’ yellow flag beside a post generally indicates that a) the user made it themselves and b) that they’re happy to share the knowledge. This creator pointed to a ControlNet workflow in comments when sharing their process.
Wait, what’s ControlNet? ControlNet is a Stable Diffusion model that allows users to copy compositions from a reference image. With many text-to-image generators, adding an image serves only to give the AI a vague reference point. ControlNet essentially recreates the parameters of the image. There’s a good more detailed explainer on ControlNet here. This is, like most gen AI developments, great news for most AI creators and terrifying news for most traditional creators. But it’s not, currently, a mainstream tool.
The above image is taken from a shared workflow for a currently popular image of a ‘checkered village’. Part of a recent trend for spirals or checks on images, facilitated by ControlNet functionality. The innovation initially focused on allowing for the hiding of QR codes in images. Something that does work!
This is the corner of the internet where people are genuinely experimenting with generative AI tools and pushing beyond relying on pop culture text prompt gimmicks (what if Luke Skywalker, but Pixar! What if Marvel, but countries! What if Good, but Bad!).
If the above screenshot looks intimidating and incomprehensible, that’s OK! I might do a deep dive on ControlNet in a future issue, but it’s mainly suited to the people who grasped tricky tools in the pre-generative AI era and now would like new tricky tools to grasp. They’re making some cool stuff, and it’s likely this is where plenty of breakthrough viral moments will come from. But these hack-y workarounds will probably find their way into future iterations of mainstream tools as more user-friendly options.
Which doesn’t mean the hidden text trick is not accessible to the masses. Glif, an AI generator website, hosts a user-friendly version of the ControlNet hidden text feature here. It allows for a single word, all caps works best, to be hidden in an image. It made this:
The tool works best with image prompts that lend themselves well to hidden letters (ie, messy bedroom > ice rink). If, like me, you need to use a longer word then go with a smaller font. But long words are trickier to distinguish (spoiler: the above image says ‘subscribe’).
The text-to-video tool Pika Labs has also got social media traction for offering the same hidden message feature, but my results there were patchy. And the gimmick, for it is a gimmick, lends itself better to a static image. If you are keen on a video then using ‘waves’ in your prompt seems to be the safest bet. So far, there haven’t been any breakthrough efforts from Midjourney or Dall E where text is still a major issue.
So what will the near future look like, when everyone gets super into this trick? It will likely break in a few different ways:
A solid advertising gimmick. They have already begun to be spotted in the wild. Some smart, some not, quickly considered hackneyed because: advertising
A boon for the conspiracy theorists! It’s no surprise that the early breakout star of this image trick had a conspiratorial message. Many of the shares were focused on global control messaging, rather than its AI tool origins. With a US election next year this could have the longest tail
A new meme factory, obviously
Now we’re beyond the initial wave of text-to-image hype the hidden messages are part of the next wave of creations designed to create engagement. 3D illusions are getting shares, the previously mentioned Spiral Town meme is picking up traction, the next mini era of AI images will likely get people thinking about fresh ways to engage with a visual.
Small bits #1: New dystopia just dropped
An apparent leaked ad for Google’s new Pixel 8 and Pixel 8 Pro phones has done numbers for the wrong reasons. The video, first shared by 91 Mobiles, demonstrates the ability to change faces using AI. If the video does officially surface it may get an edit to not appear like it is creating a hellscape where family photos exclusively feature smiles. It has, understandably, generated some horrified reactions.
Small bits #2: Authors and data training sets
On September 25, The Atlantic, already hot on AI and copyright stories, published These 183,000 Books Are Fueling the Biggest Fight in Publishing and Tech along with a handy search tool for authors to search if their work had been used to train generative AI systems. There has been a steady stream of published authors upset at seeing their work in there over the past 24 hours.