Uncle’s Diary · Learn AI Simply · 5 minutes to read · 11 seconds to watch
Last week I made my dog talk. It cost nothing, it’s eleven seconds long, and there’s a watermark in the corner.
The confession
Before you assume I’ve become technical: the tool asked me for a “source image at 1:1 aspect ratio,” and I sat there genuinely unsure whether that was a size or a shape.
So I stopped trying to work it out and asked:
“Do we have a good enough image and voice of Boomer to do a short solo video?”
That’s my entire technical contribution. One question. Everything below came back from it.
You don’t have to learn the tool.
You have to learn what to ask for.
I don’t need to know what 1:1 means. I need to say “make the picture the shape it wants” and have someone competent handle it.
The two files
The tool wants two things: a face and a voice. Everything else is its problem.
The face I had. A photo of Boomer I’d taken myself — and no idea whether it was good enough, or whether “good enough” was even something a person could judge.
So I asked. Yes, it said: face square to the camera, eyes clear, nothing crossing the muzzle. Then it made the picture the right shape for me.
The voice I did not have. I have never written a script in my life.
So it wrote one. Boomer introducing himself in that aristocratic delivery — distantly related to royalty, saddled with an Uncle who can’t find the camera button on his own telephone. I read it and laughed. That was my contribution.
Then it explained the part I’d never have thought of.
🐾 Boomer here. A lip-sync test needs a mouth that opens. Mumble at it and you have proved nothing. So note what my script is built from — Boomer. Charles. Cavalier Spaniel. Royalty. Words that make a jaw move. Not an accident of my charm, and Uncle had no notion it was even a consideration.
The finished voice ran nearly a minute. We used the first twelve seconds — not to save time, to save credits. A full minute would have spent five times as much to answer the identical question.
And the twelve seconds still has pauses in it — about two and a half seconds’ worth. But nothing longer than half a second, so the face never sits still long enough to look frozen.
Then it handed them back finished. Two files, ready to download. The photo already the right shape. The audio already cut to twelve seconds.
I didn’t crop anything. I didn’t trim anything. I never found out what format Hedra wanted, because by the time I got there it already wanted what I had.
Download, download, drag, drag. That was the hard technical part.
What happened while I made tea
It picked one question to answer: does the mouth really move with the words, or does it just wobble?
It left every setting on the defaults on purpose, so a bad result would mean the tool’s fault and not ours.
Then it told me which four things to click. I clicked them. One was wrong. It sorted that out too.
🐾 Boomer here. “Sorted that out too” is carrying an enormous amount of weight in that sentence, but I shall let it stand. Notice the ratio, reader. Uncle contributed one idle question and four mouse clicks, one of which was wrong. That is not a lowering of standards — it is the correct division of labour, and it is available to you this evening.
Then I said something useless
The video came out. I watched it and said:
“The Hedra clip looks good.”
As far as I was concerned, we were finished. My AI didn’t accept that.
Here’s the problem I had no idea I had: a video where the mouth really moves with the sound, and one where a generic mouth-flap loops on top of it, look almost identical at normal speed. Completely different products. I couldn’t have told you which one I was holding.
So without being asked, it pulled still frames from the loudest moments and lined them up:
Same clip, a fraction of a second apart. The face is moving because of the sound, not instead of it.
I didn’t ask for that check. I didn’t know it was a thing you could check. It was recommended, then done, then explained — in that order.
Only at the end was I asked whether that was my dog’s face. That bit was mine. That bit is always yours.
The honest ledger
What I did:
Wondered out loud whether it was possible. Handed over a photo I wasn’t sure about. Read a script and laughed at it. Downloaded two finished files and dragged them in. Clicked four things. Looked at a picture and said yes, that’s my dog.
What was done for me:
Chose which single question the test should answer. Judged my photo and told me why it passed. Sized it. Wrote the script — the voice, the jokes, and the words picked to make a mouth move. Decided twelve seconds was enough and cut the clip. Handed both files back ready to upload, nothing left for me to do to them. Left the settings alone on purpose so a bad result would mean something. Worked out the four clicks. Fixed the one I got wrong. Refused to accept “looks good.” Recommended checking the mouth frame by frame. Did the checking. Explained what I was looking at.
Short list, long list. That’s not modesty, that’s the arrangement.
What free actually costs
The free version stamped its logo in the corner. Most people treat that as the thing standing between them and a real result. I’d flip it round. The watermark is a receipt — proof the thing worked before any money changed hands.
The rest of the bill: free tiers are slower, and they cap quality. Mine maxed out at 720p — fine for finding out, not for finishing. And on some services, though not all, what you upload may be used to train the system.
That last one matters. Don’t test on anything private — no medical documents, no financial statements, no photographs of people who haven’t agreed. A dog’s face is fine to experiment with. Your tax return is not.
When to actually pay
When the free version already did the thing — did it, not almost did it. When you’ve hit a limit you can say out loud. And when you’ll use it more than once.
“The paid one seems better” is not a reason. It’s a feeling, and it’s the feeling that got me three subscriptions in April I have never once opened.
Your turn
Pick a tool you’ve been curious about and haven’t tried because it looked like it wanted money. Change the words in brackets and say this:
“I want to find out whether [name of tool] can [the thing you want it to do]. I don’t want to pay anything and I don’t know what the settings mean. Use the smallest test that would actually answer the question, leave everything on the defaults, and just tell me what to click. Then help me check whether the result is any good.”
You’ll either end up with a tool that works, or you’ll have saved yourself a subscription. Both count as a good evening.
Next, in I Didn’t Even Know It Was Broken — the thing that was wrong with Boomer’s recording, which I never noticed, never mentioned, and was perfectly happy to publish.
All of this was done on free tiers in one evening in July 2026. Tools change their pricing constantly — check current terms before signing up for anything.