smartechsummit.comIt’s time again to drop in on Berlin to witness IFA on this, the event’s 100th anniversary. As...
The One Tech Demo That Will Haunt Me From MWC 2024
AI has the capacity to be truly life-changing. It can potentially turn the art averse into artists, the borderline illiterate into somewhat competent writers, and one day it could even turn a budget-less amateur filmmaker into Sam Raimi. But that’s all still a while away yet. While things like AI images do find their way into the odd professional project, AI art still has tremendous entertainment potential. When it came to entertainment, one company’s on-device AI really stood out at MWC 2024.
Picture the scene: The Vatican buzzes with anticipation as Hulk Hogan, dressed in white attire, stretches out one of his “24-Inch-Pythons” toward The Pope. Clasped in his hand — the same hand that held many a WWF Title — is a cup of yogurt. Grateful, the Pontiff takes a sip from the Hulkster’s yogurt cup.
While the above scenario may have been fun, you really don’t have to imagine it. Instead you can pump it into MediaTek’s on-device image generator, and you’ll get something close to the scenario I mentioned above.
Not every device can do this; in fact, no other device I entered a prompt as wacky into came close. MediaTek had far and away the best on-device AI image generator at MWC 2024, but unfortunately, I’m not allowed to shout about it on our best-of list. A fact that will haunt me almost as much as that picture of the Hulkster shoving a cup into John Paul II’s mutated AI arm will.
It’s on par with an overactive imagination
If you want the official name of this mind-blowing AI tool, it’s the “SXDL Turbo Demo,” or the “Video Diffusion Demo” depending on where you’re viewing it. The sheer scope of the AI image generator is probably the most impressive thing about it. The fact that it generates as you’re typing, so you’ll see your weird ideas grow and change before your eyes in some kind of twisted slideshow, is the second-best thing about it. It is Stable Diffusion-based, so with the right resources and tuning you could re-create the accuracy, but the on-the-fly generation will be a big miss.
Pro wrestlers visiting the Vatican with dairy products aside, MediaTek’s model is capable of replicating an absurdly wide range of seemingly obscure requests. Another participant managed to conjure up a rendering of The Stig from “Top Gear” riding some kind of fruit around a track. It recreated that perfectly.
Some of my weirder suggestions included Frank Zappa fighting a saber duel with Kevin Spacey, an obese marmoset playing table tennis with a shoe, and perhaps most weirdly, Kevin Keegan (a former soccer player and manager) impersonating Queen guitarist Brian May.
Every other image generator used that week failed to reproduce even mildly weird requests. They’re great if you’re taking things seriously, and want to see the sun setting behind a castle on a hill or something. But go a bit weird, and request a picture of Ed Sheeran shoving Lego up his nose while watching said sunset, and you’ll just confuse most models. MediaTek’s model loves that sort of nonsense.
Like other image generators, there’s obvious room for improvement
While it is quite possibly the best model I have ever used when it comes to illustrating the weird things that fly through my brain, MediaTek’s model isn’t perfect by any means. It struggles with many things other image-generation tools have trouble with: expect extra limbs, weird and deformed arms, or multiple fingers. Sometimes it gets confused and puts a person you’ve mentioned in a weird spot. We didn’t ask it to dress Hulk Hogan like the pope, for example, he was just supposed to feed the revered religious leader some yogurt.
Although it’s great at generating obscure pop-culture pictures, the more complex you make something, the more likely it is to mess up. This was illustrated when we asked for a picture of an obese marmoset playing table tennis with a shoe. It was doing great until we typed the word shoe, then the model got confused.
We’d have accepted a shoe appearing in place of a paddle, or a shoe returning the ball from the other side of the table, but there was no shoe involved. The paddle the marmoset was using disappeared, too. Instead, it was just sat there at the edge of the table, all plump and confused.
Despite these apparent shortcomings, plenty of people were desperate to take the device running the SXDL Turbo Demo home with them, myself included. Honestly, I could type random nonsense into that thing all day.
Source: http://www.slashgear.com/1534334/one-tech-demo-that-will-haunt-me-from-mwc-2024/
You may also like
