Each week, the social media hype-train appears to seek out new methods to sensationalize generative AI instruments. Most not too long ago, a brand new method that permits customers to supply optical illusions went viral, with some describing the outcomes as AI-generated photographs with “subliminal” messages.
The method, referred to as ControlNet, primarily lets customers have extra management over the generated picture by specifying further inputs—on this case, letting you create photographs or phrases inside different photographs. Some customers characterised this as a type of “hidden message” that may very well be used to implant strategies within the type of refined visible cues, like a McDonald’s “M” logo showing within the outlines of a film poster.
For instance, here’s a viral tweet displaying a picture of cute cats that spell out the phrases “Homosexual Intercourse”:
On Twitter—which is now referred to as “X” for some purpose—algorithmically-boosted blue-check customers puffed up the method as both nefarious or revolutionary, posting varied photographs with “hidden” messages. However the messages are usually not very refined, and the strategy for creating them—whereas definitely novel—is pretty easy.
“I feel the ‘subliminal message’ angle is a bit sensationalist/tech-bro-y,” Apolinário Passos, an AI researcher at HuggingFace, advised Motherboard.
ControlNet makes use of the AI image-generating software Steady Diffusion, and one in all its preliminary makes use of was producing fancy QR codes utilizing the code as an enter picture. That concept was then taken additional, with some customers creating a workflow that lets them specify any picture or textual content as a black-and-white masks that implants itself into the generated picture—type of like an automatic, generative model of the masking software in Photoshop.
“What occurred there was that this consumer found that in the event that they used the QR Code ControlNet however as a substitute of feeding it a QR code, they fed it another black-and-white patterns, they might create good optical illusions,” mentioned Passos. “Now you can ship a conditioning picture and the mannequin blends in a sample that satisfies that whereas nonetheless making a coherent picture on the similar time.”
So whereas ControlNet-produced photographs will most likely stick round, the fact of the tech is much much less sensational than the AI hype crew makes it out to be.
“Will this be a subliminal message nightmare the place manufacturers cover their brand/merchandise into seemingly harmless photographs and manipulate us? In all probability not,” mentioned Passos. “Will we see this in advertisements very quickly, however with a double-meaning the place you look/look once more/squint to see […] and have individuals speaking about it? I feel so!”
