The avatar in the ad and the burger on the poster

When we talk about working together, I have already heard more or less this sentence: “surely I just upload a photo and five minutes later I have an avatar”. It is hard to be surprised, because that is exactly what a lot of ads for AI avatar tools look like: you upload one photo and a few minutes later (or a short recording and by the next day at the latest) you get a digital version of yourself that will record a training video, an ad, a welcome message for your website and anything else that comes to mind. The technology in question really exists and impresses me a great deal, but there is quite a gap between what you see in the ad and what you get day to day, and it is worth knowing about before you spend your first money on it.

An AI avatar from one photo in five minutes

The promise is simple, and that is exactly why it works so well: one photo, a few minutes and an avatar “for everything”. In the ad you see a figure who speaks fluently, looks straight into the camera, moves naturally, and whose lips match every word. I have read accounts from people who have tried it, and they usually go the same way: someone buys a subscription, uploads their photo, generates their first video and is disappointed, because the figure on screen does look like them, but it blinks in an odd rhythm, has stiff shoulders, or halfway through a sentence pulls a face that doesn't fit what it is saying at all. The first thought is that something was done wrong, and I think that in many cases the issue is something else: what the video in the ad actually is.

An AI avatar ad is like a burger poster

The simplest comparison that comes to mind is the poster in a fast food restaurant. The burger on the poster is tall and juicy, the lettuce looks crisp and the cheese melts exactly where it should, while what you get on your tray usually looks quite different, even though the ingredients are the same. Nobody is openly deceiving anyone here: it is simply marketing, and the poster shows the burger in the best possible version, prepared specially for the photo by people who know what they are doing, quite unlike the one that comes out of the kitchen at lunchtime with a queue at the counter.

In my view, ads for AI tools are very similar. The video in the ad is the best result anyone managed to get, not the typical result you get on the first attempt, and the difference between the two can be as big as the one between the burger on the poster and the burger on the tray.

Who is really behind the video in the ad

I suspect that behind a video like that there are usually several people, each a specialist in their field. They know the tools and where their limits are, they have many experiments behind them, they have tested a great many photos and recordings and know what source material gives the best effect (lighting, framing, clothes, background, speaking pace, even how the hands are placed). They put time into it, and often money too, and from all of that they picked what came out best for the ad.

I don't hold that against them, because anyone showing their product usually shows it at its best, and so do I. It is worth remembering, though, that when you watch a video like that, you are most likely looking at the work of a team that knows exactly what it is doing, not at five minutes of one person who uploaded the first photo from their phone.

Why AI generates something different every time

The second thing the ads don't mention matters even more to me. Even if you manage to get a decent result once, that doesn't mean you will get a similar one next time. Models that generate images, sound and video work with an element of randomness, so the same prompt can give a different result each time: once very good, once noticeably worse, and once something I wouldn't show anyone.

With images and video I can save the settings on my own hardware and reproduce the same result (with music, not even that works for me), but only for an identical prompt. With an avatar, the whole point is that the figure says something different each time: a new script means a new generation, and a new generation is a new unknown. In practice, then, an avatar “for everything” means that each new video is a new attempt, not a copy of the previous one, and in each of them something can go differently.

For a business, this means adding to the budget not just the subscription but also the time for further attempts and for choosing the best version, and that time is hard to predict in advance, because it depends on how far a given script, shot or movement strays from what the model handles well.

Generating images, video and songs at home

I see this every day myself. I generate images, short videos and music on computers that sit at my home, with smaller models, weaker than the ones the most expensive services use, but in my experience the principle is the same. Sometimes I get really nice images and videos, and sometimes something that goes straight in the bin; one thing comes out right the first time, and another won't come out despite many attempts. When I was preparing a short piece with a generated presenter, an automatic selection rated some shots as good, while in others, with the same prompt, the figure suddenly looked away from the camera or the frame jumped for no reason.

Songs are similar: I picked each one I have made so far from between eight and twenty-four versions. The first song I published on my YouTube channel I picked from twenty-four generated versions (in three different genres).

What an AI avatar really costs

I don't want to leave the impression that none of this works, because it does: good avatar videos can already be made, a voice can be cloned, and sound and short clips can be generated at a quality that seemed a long way off not long ago. Getting to a result like that takes time, effort and usually money, though, and I wouldn't expect a subscription at ten or thirty dollars a month to give you an avatar you can use for every purpose.

AI really does cut costs, but not to the bare minimum. I give these numbers only as an illustration, not as a price list: if a music video once cost, say, PLN 10,000, then today, with the help of AI, it might cost around PLN 2,000, that is, several times less, but still not a few dozen or a hundred złoty, as the price of a subscription might suggest. The gap between the subscription and the cost of the finished video is filled by someone's work: preparing the source material, further attempts, choosing, fixing and putting the whole thing together.

New skills: someone who directs the AI instead of an editor

In my view, AI largely shifts skills rather than removing them. Perhaps you no longer need a video editor, because AI can do part of their work, but you do need someone who will steer the AI exactly where you want to go: someone who knows what source material to prepare, how many versions to generate, how to tell which one is good, and when to stop fixing. That person doesn't have to be able to cut a video by hand, but they do have to know what the end result should look like and how to get there with a tool that answers a little differently every time.

For me, this is where the real change lies. The cost of simply generating an image, a voice or a few seconds of video has fallen a great deal, while the cost of bringing that to a result you can comfortably show an audience has fallen much less, because it is largely the cost of someone's knowledge and time.

I have gathered more misunderstandings like this about AI on the homepage, under What's worth knowing before we talk.

In short

• The video in an ad for an avatar tool is usually the best result anyone managed to get, not the typical outcome of five minutes with one photo.
• Models that generate images, sound and video work with an element of randomness, so one good result doesn't guarantee the next one will be similar.
• At home, good and bad generations alternate with the same prompt, and I picked the first song I published from twenty-four versions.
• AI clearly cuts the cost of video and audio material, but not down to the price of a subscription.
• In my view the skills change: instead of someone who edits by hand, you rather need someone who steers the AI towards a specific result.