Content
Yes, it's AI. The difference is who's directing it.
An engine doesn't make a piece: it makes a clip. What decides whether it's any good is judgement — which model survives the shot you actually need, which voice sounds like a person instead of a narrator, and what's better left off screen. Everything on this page is live or was made for a real campaign.
The test
Four models, one photo, one prompt.
Video generators got very good, right up until you ask them for fine detail on something that moves fast. A mouth mid-laugh is the worst case: teeth appear and disappear, and the model has to invent the inside of the mouth in every frame. Add braces — a metal structure with its own geometry, one piece per tooth — and half of them give up. In an orthodontics campaign the smile is the product: one broken frame kills the clip. I couldn't find a public test that measured this, so I ran one.
Braces hold in every frame, and the face stays the same person from start to finish.
USD 0,61 · 4 s
Braces fine. It tends to duck the head while laughing, which you can fix by asking.
USD 0,67 · 4 s
Braces fine, and the most expressive movement of the four.
USD 0,78 · 4 s
The braces smear right at the peak of the laugh. It's the cheapest of the four, which is exactly why it's worth knowing beforehand.
USD 0,39 · 5 s
First generation from each model, no editing and no picking the best attempt. Same starting image, same prompt, same resolution. The whole test cost under USD 2.50.
How each frame was checkedWhy the second piece is cheap
The character, the voice and the brief are built once.
The first piece carries the cost: finding the face, picking the voice out of several, locking how the subtitle looks, deciding what can and can't be said. All three get saved, and every new video reuses them. That's why a four-piece campaign doesn't cost four times the first one.
The character
Before the video there's a photo. The face is generated once as a still — the one on the left — and the video comes off it: the engine animates that image, it doesn't invent a new one. The file gets saved, so every new shot in the campaign starts from the same face. The place, the script and the framing change; the person doesn't. And she doesn't exist.
Dar+ Sonrisas · braces campaign · 22 s
The voice
You pick it by ear, with the same text read by every candidate. The winner gets locked with its exact configuration — voice, seed and settings — and the chosen take is the asset: it never gets regenerated, because next time it would come out different. Listen to all four reading the same line.
«Me pasé cuatro años diciendo: el año que viene empiezo con los brackets».
Four of the six candidates tested for this campaign, reading exactly the same text. The takes are in Spanish because that's the campaign's market.
The brief
The brand's typeface, its colour, the subtitle that highlights the word being spoken, and the rules about what can't be shown. It runs off code rather than by eye: change the text and everything else stays exactly where it was.
This one came back with the voice stumbling over a phrase — it sounded like a stutter. It wasn't the configuration, it was the text: a pair of quotation marks the voice read as a change of tone. The text was fixed and it went back up the same day.
The work
Three pieces, three different problems.
Three formats for three clients. Each one has an open log with the whole process: what got dropped, what the client sent back, and why the surviving version won.
Person to camera
Someone telling you something first-hand, with the subtitle following the voice word by word. It's the format that performs best with a cold audience and the easiest one to ruin: if the voice doesn't sound human, you can tell within two seconds.
Cliencer · 37 s
see the logThe garment, with nobody in it
The owner doesn't want a model on the homepage — with a person in the shot, the clothes stop being what you look at. Each garment from the real catalogue was animated from its product photo as a museum piece: black background, one spotlight, the fabric breathing. No people and no voice.
Auba Oversize Style · 5 s · proposed homepage video, awaiting the owner's sign-off
aubaoversizestyle.comAn ad with the real product inside it
What's generated is the world around it; the app screen is the real one, captured from the product. This is version 6 — the five before it are in the log, each with the reason it changed.
Cliencer · Spain campaign · 30 s
see the logThere's more in the logs: the three different openings of the same video that were tested to pick one, the original music generated for a run of stories, and the wave of scripts currently in production. All of it lives in the Dar+ Sonrisas log, written as it happens.
The limits
What I won't do with this.
This isn't a list of principles: it's what producing taught me, and every point has a date in some log.
First-person testimonials with a generated face and voice. An "I did this myself" from someone who doesn't exist gets spotted, and the backlash lands on the client's brand. You can imply it without lying: the person who lived it is always a third party.
Treatment results presented as if they belonged to a real patient. In healthcare that's exactly what gets scrutinised, and an image can't be walked back with a caption underneath.
Hiding that it's AI. The label goes on the piece. What's being sold is that it's well directed, not that it fools anyone.
Your business data in a public piece. The method gets shown; the numbers stay in.
How it works
Four steps, and the first two are your call.
The order isn't arbitrary: the cheap things get approved before the expensive ones. A script is rewritten in minutes; a video shot the wrong way has already been paid for.
Concept and script
What the piece says and in which exact words. This is where you decide what your customer is really weighing at the moment they choose. You sign it off before a cent goes into imagery.
The voice
Several candidates read your script and you pick by ear. It's the cheapest step and the one where judgement shows most: an unconvincing voice sinks a video that cost twenty times more.
Image and video
Only now does the expensive engine come in. The character gets locked first, the shots come off it, and the model is chosen for what that particular shot has to survive — not out of habit.
Edit and check
Subtitles from code, music where it earns its place, and a frame-by-frame pass hunting for odd hands, broken textures and jumps. A human does the checking: a video can look fine at full speed and still hide three broken frames.
What you get
- The finished video in vertical, ready to post or to run as an ad.
- The script and the voice as separate files, to reuse wherever you like.
- The character and the brief, locked: the next piece starts from there.
- The variants to test, and the log recording what was tested and how it went.
Bring me something you want to show.
A product, a service, an idea that today takes a paragraph to explain and would land better in thirty seconds. The first conversation is about whether it's worth doing at all, not about selling you a campaign.