Why a business needs a digital presenter
The task is usually the same: you need a person in frame explaining something, but there is nobody to film, no time or no budget.
An online school needs a teacher for lessons. A shop needs a host for a product review. A company needs a presenter for internal instructions, of which there are a hundred, and nobody is going to reshoot them every time they are updated. In all these cases a synthetic presenter saves not so much money as reshoot time: the text changed, you regenerate the clip in ten minutes.
That is where the real value lies, not in «replacing live people».
Step 1. Get a face
There are two paths, and they are fundamentally different.
Path A: generate a person who does not exist
You generate a face in an image model. It is the most predictable option: the person does not exist, so moderation filters do not pick on it and generation will not break down in the middle of a project.
What matters at this step:
- The face is large and in focus. The model will read the features from this image. A small face in means mush out.
- Neutral light and a neutral expression. Hard shadows and a broad smile can stick to the character forever.
- Collect several angles. Front, three-quarter, profile. As soon as the presenter turns their head, the model has to invent what it has not seen.
How to build a full set of angles is described in detail in the article on character consistency. The prices are laughable: an image in Nano Banana Pro costs about 25 rubles, in GPT-5 Image about 46 rubles, in Nano Banana 2 about 13 rubles. A set of five angles comes to a little over a hundred rubles. Do not skimp here: a mistake at this step will ruin every generation that follows.
Path B: take a photo of a real person
It works better, because a live face is always more convincing than a generated one. Only one thing matters here: use a person’s face with their written consent, otherwise it is no longer an avatar but a deepfake with all the consequences.
But there is also a purely technical problem that almost nobody warns you about in advance.
Moderation: some models will simply refuse
You upload a photo of an employee, write a prompt, press generate and get a refusal. Not an error but precisely a refusal: the model does not want to work with this request.
It happens because model developers fear deepfakes more than you fear wasting money. So they build in filters that trigger on realistic photos of people, especially if the prompt involves speech. The model cannot verify whether you have consent, so it refuses everyone just in case.
The problem is that developers have different policies, and nowhere is this written down in plain language in advance.
A telling example: Seedance 2.0’s description says directly that the model is not suitable for photos of real people. It holds a character between shots very well, but on a real face you will hit the filter. Meanwhile the very same task may pass calmly in another model.
What to do after a refusal
First: do not try to trick the filter. Rewording along the lines of «this is not a real person, honestly» does not work and can get the account blocked. There is no point in it.
Second: switch models. It is the simplest solution and usually the right one. If Seedance refused on a real face, try Veo or Kling: they have different policies. That is exactly why access to several models in one window saves not money but nerves: you do not need five accounts to find out which one will agree.
Third: change the source. If the task does not require a specific person (and most business tasks do not), generate a face that does not exist. There will be no refusals at all: a synthetic character has no rights that could be violated, and filters do not pick on it.
Fourth: remove triggers from the prompt. Sometimes the refusal is caused not by the photo but by the wording. Words about celebrities, politics, specific names and requests to «speak on behalf of» someone raise the chance of the filter triggering.
The practical takeaway
If you need an avatar at volume and predictability matters, build the process on a generated face. It is always available, never refuses, never quits and never demands a full reshoot. Keep a real face for cases where recognising the person is genuinely part of the value.
Step 2. Animate it and make it talk
Now you have a face. It needs to be animated and synced with speech.
For this you need a model with native sound: it generates video and audio together, so the lips match the speech. This is not the same as laying a voice-over on top of a finished video: with an overlay the articulation will not match, and it will be noticeable from the first second.
KADR prices, 6-second clip, 1080p, sound on, October 2026:
| Model | Price, rubles |
|---|---|
| Google Veo 3.1 Lite | 84 |
| Google Veo 3.1 Fast | 126 |
| Kling v3.0 | 168 |
| Google Veo 3.1 | 417 |
This is not a public offer.
If you already have a recorded voice track, there is another route: OmniHuman 1.5 makes the person in a photo speak or sing exactly to your audio (2 to 30 seconds, 720p). 10 seconds cost 195 tokens (about 308 rubles).
Note: sound is mandatory here, and it doubles the price. More on what makes up the price of a generation is in the article on how much a minute of AI video costs. The same clip on Veo 3.1 without sound would cost 132 tokens instead of 264. For an avatar, sound is the whole point, so you cannot save on it. But you can save on resolution: run drafts at 720p.
The problem with non-English articulation
You have to be honest here, because almost nobody writes about it.
The models were trained mostly on English speech. Their English articulation is noticeably better than their Russian. In Russian the lips move similarly but not precisely: they hit the rhythm, but not always the specific sounds. An attentive viewer notices this even if they cannot say exactly what is wrong.
The practical takeaway: run a test before you build the avatar into a project. Take one phrase from your real text, generate it in two or three models and watch the lips. It costs a few hundred rubles and saves you from disappointment at the stage when it is too late to redo things.
What helps:
- Short phrases. The longer the line, the more likely the articulation drifts towards the end.
- A medium shot instead of a close-up. In a close-up every inaccuracy is visible. In a medium shot the viewer’s brain fills in the rest.
- Calm speech. Fast delivery and emotional bursts are harder for the models.
How to make a long video with a presenter
Models generate segments of a few seconds, while a lesson or an instruction lasts minutes. So the video is assembled from pieces.
The scheme:
- Split the text into lines of 8 to 15 seconds.
- Generate each one separately, feeding the same presenter image as a reference.
- Stitch them in an editing program.
- To keep the joins from standing out, cut away to other shots: a slide, a screenshot, a title. It is a standard editing technique, and it also hides small differences in appearance between scenes.
If the presenter must be continuous, use the last-frame technique: the final frame of one scene is fed as the starting frame of the next. But remember that errors accumulate this way, and every few scenes you need to go back to the original image.
Where an avatar does not work
An honest list, so you do not waste money.
- Where trust in a specific person is needed. A doctor, a lawyer, an expert. A synthetic presenter in that role, if recognised, wipes out trust entirely.
- Where a real product in hand is needed. An avatar will not hold your product: the model will draw a similar one, not yours.
- In long emotional monologues. The longer the viewer looks at the face, the more likely they sense something is off.
An avatar is good where it plays a service role: instructions, announcements, short explanations, repetitive content that is updated often.
In short
An AI avatar is two steps: fix the face as an image, then animate it with a native-sound model. The face costs pennies to generate, video with sound costs twice as much as without, and non-English articulation needs a test before work starts.
And the main practical advice: if you need an avatar at volume, build it on a generated face. A real photo will hit moderation filters where you least expect it, while a synthetic presenter always works.
