Everyone is talking about AI models. Hardly anyone talks about the hardware they run on, and that is where the change has been just as big.
We had a video on our home page explaining snapWONDERS Vaultify. It did the job, but it was 40 seconds of feature labels at 372×252 pixels, and it looked its age. So on Sunday 18 October we tried something. Could we make a better one with AI, and what would it actually take?
What we built
A 17-second commercial. A family dinner. Someone takes photos of her grandparents. Those photos go into Vaultify and each one hides inside an ordinary holiday photo, a beach, a rice terrace, a waterfall. Later an aunt scrolls through the phone and sees only scenery. At the end, one tap brings the grandparents back.
Every person in it is AI-generated. So are the photos, the movement, the voice and the music. We wrote the story, picked every shot and made the calls. The machines did the rest.
What it took
All of it ran on one modern AI GPU, an NVIDIA L40S with 48 GB of memory, between 12:09 pm and 5:30 pm on a Sunday. That time includes all my review rounds.
Job Modern AI GPU (measured, approx.) Standard office desktop (estimate) 64 candidate photos about 1 second each about 10 to 20 minutes each 5 video clips, 5 seconds each about 8 minutes each would not run in 16 GB of memory; with more, about 1 to 3 days per clip 4 music takes about 1 minute each about 1 to 2 hours each Voiceover seconds seconds Whole video about 5 hours, under A$10 2 to 3 weeks non-stop, if at all
By “standard office desktop” I mean a normal work PC. Core i5 class, 16 GB of RAM, no graphics card. That is what most of us still sit in front of.
The old way would be casting about eight people, a location, a crew, a voice artist, licensed music and an editor. A rough industry ballpark is two to four weeks and A$10,000 to A$50,000 or more.
Every model we used is free and open-licensed (Apache-2.0): FLUX.2 [klein] 4B for the photos, Wan 2.2 for the movement, Kokoro for the voice and ACE-Step for the music. Claude did the orchestration, wrote the scripts and drove the models.
What went wrong
Plenty, and that is the useful part.
- Phone screens. The video model cannot draw a phone screen. Every time, it came out as garbled nonsense. In the end we kept screens out of the AI shots and drew the phone screens ourselves.
- The cast. I asked for a Filipino family. The first batch came back with a mix of faces that did not belong at the same dinner. It took better prompts to get a consistent family.
- The story. The first version looked great and the music was right, but I watched it and could not tell what had happened. Family dinner, then beach photos, then a logo. So we rebuilt the middle to show each photo sinking into its own beach photo, right on the music’s peak.
- A second opinion. We asked another AI model to review it cold. It spotted that the “scenery only” gallery still had two food photos in it, and that we never showed anyone getting their photos back. Both fixed.
- Memory. The first video render ran the GPU out of memory at the very last step. Even a 48 GB card has limits.
Captions are on
Captions are on by default. Accessibility is important to me, and I don’t think it should be an option you have to go looking for to turn it on, but rather to turn it off.
So the captions were part of the brief from the start, not something added at the end. Plenty of people watch with the sound off, and plenty can’t hear it at all. Everyone should be able to follow the story.
Captions also took the least compute of anything in the project. They are text.
Where this leaves us
The models get the headlines. But a few years ago, this project would have needed a film crew or weeks of compute. On Sunday it needed one GPU, an afternoon and someone who knew what the story had to say.
The AI still cannot decide what the video should mean. It could not tell that the first version did not explain anything. That part is still ours.
Pretty happy with where it landed. You can try snapWONDERS Vaultify with your own photos, and let me know what you think of the video.