If someone offers to make an AI video recreation of your wedding, just say no. This is the tough lesson I learned when I started trying to recreate memories with Google’s Gemini Veo model. What started off as a fun exercise ended in disgust.
I grew up in the era before digital capture. We took photos and videos, but most were squirreled away in boxes that we only dragged out for special occasions. Things like the birth of my children and their earliest years were caught on film and 8mm videotape.
When I got married in 1991, we didn’t even have a videographer (mostly a cost issue), so the record of that date is entirely in analog photos.
In general, there was no social media to capture and share my life’s surprising moments; I can’t point someone to Instagram, Facebook, or TikTok and say, “Don’t believe me? Check out this link.”
I do have a decent memory, though, and I wondered if I could combine it with a little AI magic to bring these moments to life.
Memory machine
Recently, I signed up for the three-month Google Vertex AI Studio trial, which includes access to 300 Veo credits. Veo 3 is the remarkable Gemini model that can produce synced audio and video with one prompt.
For my test, I chose a couple of memorable moments from my early career and my 20s in Manhattan. These are 100% true stories that happened to me, but I have no visual record of them.
For the first one, I described a young, skinny, bespectacled man with curly hair (yes, I once had a full head of curly hair) meeting a famous and Tony Award-winning comedian in Times Square on Broadway. The comedian was Jackie Mason (ask your grandparents), and I wanted his autograph. He stopped, but as I spoke to him and he inexplicably starting quizzing me about which TV to buy, a pigeon pooped on my head. Mason didn’t notice, I kept my composure, and answered.
For the prompt, I painted the scene in broad strokes, describing my business attire, the year – 1989 – and how Mason looked with his curly hair and “cherubic face.” I included the bit of dialogue I could remember, and the action of me touching my head and realizing what happened. Then I fed Veo 3 the prompt.
A few minutes later, I had a decent recreation of the scene, complete with the pigeon. The guy didn’t look that much like me, and the Jackie Mason character bears only a passing resemblance to the once-iconic comedian.
Still, I was emboldened, and searched my memory for another memorable moment from my 20s.
I settled on the time I tried to impress my first boss with my tech skills. His laser printer (yes, kids, they existed in the 1980s) was running low on toner, but I remembered that you could extend the life of a cartridge by removing it from the printer and shaking it. So, that’s what I did, but the cartridge panel was stuck open and I proceeded to shower myself and the office with black toner as my stunned boss looked on.
In my prompt, I described the scene, including the wood-panel walls of the circa 1986 office, and included a brief description of myself and my bald, middle-aged boss who was seated at his desk. The dialogue included me explaining what I could do, saying, “Sorry,” and my boss’s good-natured laugh.
The results this time were even better. Even though neither character looked like their real-world counterparts, the printer, desk, and office were all eerily close to my memory, and the moment when the toner went everywhere was well done.
If I could open my brain and show people my memory of that moment, it might look a little like this. Impressive.
A union too far
Imagining a lifetime of memories rebuilt with AI, I wracked my brain for another core recollection. Then it hit me: my wedding.
It has always bothered us, particularly my wife, that we didn’t have a wedding video. What if I could create one with AI (I know, I know, the foreshadowing is too heavy).
It would not be enough to simply describe a wedding in Veo 3 and get an AI wedding video featuring people who looked nothing like us. I also knew, however, that you could guide an AI with source material. I have a lot of 34-year-old wedding photos. I grabbed a scanned image of one that featured me and my wife shortly after the ceremony, walking hand-in-hand back down the aisle. I liked the image not only because we were clearly represented but also because it featured some of our wedding party and guests.
These are worse than false memories; it’s an active distortion of one of the most important moments in my life.
With the hope of creating a long-sought-after wedding montage (of just eight seconds in duration), I crafted this prompt.
“I need a wedding video montage based on this wedding photo. The video should look like it was shot on HD-quality VHS tape and feature 2 seconds of the ceremony, 2 seconds of everyone dancing, a second of the groom feeding the bride wedding cake, a second of the bride throwing the bouquet, a second of the newlyweds leaving in a limo as everyone waves goodbye.”
Ambitious, I know, but I thought that by giving the model specifics on scene duration, it might squeeze it all in.
Instantly, I hit a speed bump; my Veo 3 Trial didn’t allow me to include a source image. If I wanted to start with a photo, I’d have to step back to Veo 2, which also meant I’d lose audio. That wouldn’t be a big deal, though, because, as described in the prompt, there really isn’t that much dialogue.
It took another few minutes for Veo 2 to spit out a few videos. All of them start with the base image, but to put it plainly, they are very, very wrong.
In each video, the thread of consistency snaps almost instantly, and my wife and I transform into other people. At one point, I’m dancing while holding a cake, and in another, my wife doesn’t know how to let go of the bouquet she’s supposed to throw. We awkwardly feed each other cake and sort of dance together.
The video is horrifying because it looks kind of right but also very wrong. These are worse than false memories; it’s an active distortion of one of the most important moments in my life. I showed the videos to my wife, who was appalled and told me it would give her nightmares.
It was hard to disagree, but I did remind her that the models would improve and a future result would be better. She was unmoved, and looked at me like I had sold one of our children.
What I did is no different from people reanimating photos of dead relatives with My Heritage. Whatever the image starts with, everything after that first millisecond is false, or worse, it’s memory corruption. If you spent any time with that person when they were alive, that’s the true memory. An AI creation is guesswork, and even if it’s good, it’s also fake. They never moved just like that at that specific moment.
In the case of my wedding memories, I realize they’re better left to the gray-matter movie projector in my head.
As for the Veo 3 creations of my other memories, there’s no base image to corrupt. The AI is not recreating my memories as much as it’s become a storytelling tool, another way to illustrate a funny anecdote. That person isn’t me, that man isn’t my old boss, and that’s not Jackie Mason, but you get the gist of the stories. And for that, AI serves its purpose.
You might also like
https://cdn.mos.cms.futurecdn.net/c3ZKbpFHX3tmZ28W2mTipi.jpg
Source link
lance.ulanoff@futurenet.com (Lance Ulanoff)
