I recorded a 30-second script at home while I was off sick, in bad lighting and a hoodie, and it took a few fluffed takes to get through it. That left me with 2 minutes 30 of footage. A couple of hours later it had become a Meta-ready ad, set in an office, with me in a black T-shirt and captions and B-roll on top. It was still recognisably me, with my own face and voice. That’s what AI video ads can do now. Below is the whole before and after, step by step, including the five-second test that stops you wasting money.

Why video comes after statics

Video is the most expensive thing you make, so it comes last. You use cheap static ads to find the message that works, and only once you’ve found it do you spend on video to push it harder. That’s the logic of the AI ad creative system this workflow belongs to: Test for pennies, then spend where the results have earned it.

So by the time you sit down to make a video, you already know roughly what it should say. What’s left is production, which means turning a rough recording into something you’d actually put money behind.

From a phone recording to a rough cut

You don’t need a studio to record the script. I filmed mine at home, badly, and needed a few attempts, so I ended up with 2 minutes 30 of raw footage for a 30-second script, with long pauses between the takes.

ChatGPT Work turns that into a rough cut. I dropped the video in and asked it to transcribe the footage and cut it down, and it stripped out the pauses and the fluffed takes to give me a first cut. That cut wasn’t quite right, so I gave it notes in plain English: Clean up the bit around the 33-second mark, and tighten that gap. The next version was my final cut, and I never had to scrub through a timeline myself.

Restyle it: office, lighting, wardrobe

This is the step that still surprises me. I sent the final cut to Higgsfield with a short brief: Change my home background so it looks like an office, fix the lighting, and put me in a plain black T-shirt. (Higgsfield sponsored the video this article is based on. The results here are the ones I got.)

What came back looked like a different shoot. The room, the lighting and my clothes had all changed, while my face and voice stayed exactly as they were, so it was a better-dressed version of me in a better room. Higgsfield then finished the job by adding captions and animated cards, and closing on a clean 9x end card.

Test five seconds before you commit

This is the step that saves the most money. Before Higgsfield rendered the whole video, it tried the treatment on a five-second clip, using two video models with very different prices.

That test showed exactly why it’s worth doing. The cheaper model couldn’t keep the character consistent, so it swapped me out for someone else entirely. He was perfectly handsome, and no use at all in an ad that’s meant to show me. The more expensive model, Seedance 2.5, kept my face and voice and simply applied the new setting. Seedance uses a reference system to hold a character steady across a clip, although even it has its flaws.

The test cost almost nothing, and it saved me from rendering 40 seconds on the wrong model. It’s also why Higgsfield’s choice of models is worth having: You can try a few, keep the one that holds up, and find that out cheaply.

Captions, cards and B-roll

Once the model was chosen, the finished ad came together quickly. Captions and animated cards went on automatically, and those two things do most of the work of making a social video ad watchable with the sound off.

B-roll adds the rest. My script has a line about needing zero coding, and I wanted a shot of someone battling with a computer to sit behind it. So I asked Higgsfield for three seconds of exactly that. Cutaways like this are small, but they’re what make a talking head look produced. The finished ad then goes straight into a campaign through the Meta Ads MCP, the same connector the rest of the system runs on.

What it costs, compared with an editor

Here are the real numbers from the build, so you can judge for yourself. On Seedance 2.5, video costs about $3.23 per five seconds, so my 40-second edited ad came to $25.80, and the three seconds of B-roll added $2.08. Those figures come from Higgsfield’s pricing, so check the current rates before you budget.

Now compare that with hiring a human editor to swap you out of a hoodie, put you in a black T-shirt and rebuild the room behind you. That job would cost a lot more than $26, and it would rarely come back the same afternoon. If you film somewhere decent in the first place, you can skip the restyle and use AI only for captions and B-roll, which is cheaper still. And if you’d rather stay off camera altogether, you can build the whole ad from generated footage.

Consent and disclosure

Get this right before you publish. Using AI to change how a real person looks or sounds in an ad is increasingly regulated. You need that person’s consent, and depending on where the ad runs, you may also have to disclose that it’s AI-generated. The rules around AI faces and voices in advertising have tightened, and in the US the FTC’s endorsement guidance is the baseline to know. My case was simple, because it was my own face and I’d agreed to it. If you’re editing someone else, get their consent in writing, and check what disclosure your platforms and jurisdictions require.

Where to start

Record one rough clip of a message you already know works, run it through the rough cut, and test five seconds before you render anything longer. That one test is the habit that keeps AI video ads cheap.

The whole workflow, including the prompts and skills for the cut, the restyle and the captions, is free in the 9x community, where our operators share what they actually run. Grab it and turn one recording into an ad.

Common questions

Can I make a video ad from a phone recording? Yes. You record a rough clip, have ChatGPT Work transcribe it and cut out the pauses, then send it to Higgsfield to restyle the setting, fix the lighting and wardrobe, and add captions and B-roll. What you get back is a finished video ad with your real face and voice.

Which AI video model keeps your likeness? In my build, Seedance 2.5 kept the character consistent, with my real face and voice, while a cheaper model swapped me for someone else. No model is perfect, which is why you test a five-second clip on your chosen model before you render the full video.

How much do AI video ads cost? In this example, video ran at about $3.23 per five seconds on Seedance 2.5, so a 40-second ad came to $25.80, plus $2.08 for three seconds of B-roll. Rates change, so check the current pricing, but it’s a fraction of what a human editor would charge for the same restyle.

Is it legal to use AI to change someone’s appearance in an ad? Yes, with care. Using an AI-altered or AI-generated version of a real person’s face or voice in an ad needs that person’s consent, and some places require you to disclose that the content is AI-generated. Get consent in writing, and check the FTC’s endorsement guidance and your local rules before you publish.