Let me start at the end: I now have a folder called ad_variants_DO_NOT_DELETE sitting on my desktop with 47 files in it. My client approved 11 of them. I didn't plan any of this.
The Part Where I Explain How I Got Here
It was a Tuesday. I was at my usual spot — that slightly-too-loud independent coffee shop near the canal where the WiFi password is written on a chalkboard in chalk that's been there since 2023. I had 45 minutes between calls. I was supposed to be eating a sandwich.
I wasn't eating the sandwich.
Instead, I was poking around an AI ad creative generator I'd been putting off testing for two weeks. The brief was simple enough: a DTC skincare brand needed creatives for a retargeting campaign. Three platforms, two audience segments, one very opinionated founder who had "a vision."
I'd been doing this kind of work manually for about three years. You know the drill — brief, moodboard, copywriter ping-pong, three rounds of revisions, someone's cousin has feedback. The usual.
The Accidental Experiment
Here's where it gets embarrassing (in a productive way, I promise).
I was trying to generate a single static ad for the 9:16 format. I misread the interface. Instead of selecting "generate 1 variant," I somehow had it set to batch mode — I think I'd been clicking around too fast — and hit confirm while reaching for my coffee.
The tool started generating. And kept generating. And then kept generating some more.
By the time I looked back at the screen, it had produced 12 variants of the same ad — different headline placements, different CTA button positions, different background color temperatures, same core visual. All from one prompt. All in about 90 seconds.
I sat there for a moment. Then I ate some of the sandwich.
What I Actually Learned About AI Avatar Video Ads (The Hard Way)
Here's the thing about AI avatar video ads that nobody really talks about in the breathless "10x your output!" posts: the value isn't in the single output. It's in the variance.
When I was generating manually — or even using AI tools one-at-a-time — I was unconsciously anchoring to my first instinct. First layout. First color choice. First CTA phrasing. I'd iterate, sure, but always from that anchor point. My "variations" were basically the same ad wearing slightly different shoes.
What the accidental batch generation showed me was that the tool had no such anchor. It was genuinely exploring the space. Some outputs were terrible (one had the CTA button in the lower-left corner, which, why). Some were fine. But three of them were genuinely surprising — layouts I wouldn't have arrived at through my normal process.
The founder, who had "a vision," picked two of those three as her top choices.
(I did not tell her they were accidents.)
The A/B Testing Angle Nobody Warned Me About
So I leaned into it. For the next two weeks, I started using the AI ad creative generator specifically as an A/B testing variant machine rather than a "make me an ad" machine. Subtle but important reframe.
The workflow looked roughly like this:
- Write one solid creative brief with locked brand elements (colors, logo safe zones, tone)
- Generate 8–12 variants per format in batch
- Do a fast cull — remove anything with obvious layout failures or off-brand color drift
- Present 3–4 survivors per format to the client
- Let the actual ad platform data do the rest
The result? We went from "one creative per format, revised three times" to "four creatives per format, tested in market, optimized based on real CTR data." Same budget. Same timeline. More signal.
The part I didn't expect: the client conversation changed. Instead of defending creative choices ("why is the button blue?"), I was presenting options and framing it as a testing hypothesis. Clients love feeling like scientists. File that one away.
Where It Still Falls Apart
I want to be honest here because the "AI replaced my entire workflow" narrative is exhausting and also not true.
Brand color consistency remains genuinely annoying. Even with locked hex codes in the prompt, the tool occasionally drifts — a slightly warmer white here, a slightly more saturated accent there. Not catastrophic, but enough that I do a manual color-check pass on everything before it goes to the client. That's maybe 10 minutes per batch, which is fine, but it's not zero.
Copy-visual tone matching is the other one. The tool is good at layout logic. It's less good at understanding that a brand with a dry, minimalist voice probably shouldn't have a CTA that reads "GLOW UP NOW ✨" even if the visual composition is technically sound. I've started writing CTA copy separately and injecting it rather than letting the generator suggest it.
And the avatar video side of things — when I started experimenting with animated avatar variants rather than static ads — the consistency issues compound. Lip sync on the avatar was fine for short clips (under 15 seconds). Anything longer and there's a subtle drift that I can't quite describe except to say it looks like someone who's had one too many coffees and is trying very hard to appear normal. (I say this with affection. I also look like this sometimes.)
The Folder on My Desktop
Back to those 47 files.
After the skincare campaign, I started doing this for every project — generating more than I needed, culling fast, testing what survived. The DO_NOT_DELETE folder is actually an archive of every variant I've generated across four client projects. I keep it because occasionally I'll be working on something new and think "wait, didn't I accidentally solve this layout problem three months ago?" and I'll go digging.
It's become a weird kind of creative library. One that I built mostly by clicking the wrong button.
I've since tried a few tools in this space. The one I've stuck with for avatar video ad variants specifically is Nextify.ai — mostly because the batch generation interface is explicit rather than something you stumble into by accident, which is, you know, an improvement on my original workflow.
The Part I'm Still Thinking About
Here's what I haven't fully resolved: if the best creative decisions I made in Q1 came from outputs I didn't deliberately design — if the variants that performed best were the ones the tool generated while I was reaching for a sandwich — what exactly was my contribution?
I wrote the brief. I did the cull. I framed the client conversation. I caught the color drift. I fixed the CTA copy.
That's still something. I think.
But honestly — if the tool is generating the variance and the market is selecting the winners, am I a creative director or just a very expensive filter?
Top comments (0)