AI ad generator: what each kind actually makes

Search for an AI ad generator and you get a category that does not exist. Under one phrase sit at least four unrelated products: tools that make a still image, tools that write the words, tools that build a talking presenter from a script, and tools that move a real performance onto a different face. They are bought for different jobs and they fail in different ways. This guide separates them, so that you can work out which one your next campaign actually needs.

The four things sold as ad generation

Ask what a tool generates, not what it promises. The answer is always one of four things.

The image. A product photo placed on a new background, resized for a placement, or reimagined as a variation. This is the largest slice of the category by volume.

The words. Headlines, primary text, hooks and variations of all three.

The presenter. A person on screen who delivers your message and was never filmed.

The whole scene. A text prompt in, an invented clip out, including people, rooms and motion.

A tool that markets itself as a full ad generator is usually chaining two of these together. That is useful to know before you pay for a subscription that overlaps with something you already have.

Generating the image, usually inside the ad platform

The quiet fact about image generation is that the ad platforms ship it themselves. Meta's Ads Manager can generate image variations from an ad's original creative, and it can generate text variations for the headline and the primary text. If your ad is a product photo with a caption, a large part of the tooling you might go shopping for is already sitting inside the account you buy media with.

What that tooling does not do is produce a person talking to camera about your offer. Generated backgrounds and cropped variations are formatting work. They make an existing asset fit more placements. They do not create the asset.

Generating the words

Copy generation is the least controversial part of the category and the easiest to misuse. A model will happily produce forty hooks for an offer it has never seen, and thirty-five of them will be interchangeable.

The useful pattern is narrow. Give the model the real specifics, the objection a customer raised last week, the exact price, the thing the product does that a competitor's does not, and ask for variations of one angle rather than forty angles. The bottleneck in paid social is rarely the number of hooks. It is having something true and specific to say in them.

Generating the presenter: the three video techniques

Video is where the differences get expensive, because three unrelated techniques are sold under one label.

Script-to-avatar. You type a script, pick a synthetic presenter and the tool produces a voice and lip movement to match. No camera, no recording, the fastest path from a document to a video. The trade-off is the performance: the pauses, the emphasis, the gestures, the small timing decisions that make a person look like they mean what they say are all invented by a model.

Performance capture. You record the take yourself on a phone, and a model transfers that performance onto a different presenter. Runway's Act-Two is the reference model: a driving performance video supplies the movement, expressions, audio and gestures, and a character image supplies the face. What you get back is your own delivery on someone else's face.

Full generation. A text prompt produces the entire clip. Flexible for scenery, unreliable for a specific product in a specific hand, and the same presenter rarely survives two prompts, which makes a series of variations hard to hold together.

Which one fits which job

A short decision list, in the order the questions usually arrive.

1. You have an asset and need it in more sizes and placements. Use the generative tools inside your ad platform. Do not buy a second product for this.
2. You need volume of copy against one angle. A language model is fine, provided you feed it specifics rather than a category name.
3. You need a presenter and have no time to film. Script-to-avatar gets you a video today, with delivery as the weak point.
4. You need a presenter and can spend ten minutes recording. Performance capture keeps your voice, timing and gestures, so the delivery is a real one.
5. You need scenery, product shots in motion or b-roll with no dialogue. Full generation is the right tool for exactly that and a poor one for a spokesperson.

None of the four repairs a weak offer. Generation lowers the cost of testing a good one, which is a smaller claim and a more useful one.

Disclosure is now part of the workflow

Three obligations sit on top of whatever you generate, and they are not optional extras to be handled later.

A presenter is not a customer. The FTC's Endorsement Guides require an endorsement to reflect the endorser's honest opinion and any material connection to be disclosed. A generated face has no opinion to reflect. Write presentation and demonstration, not testimony.

Platforms require labels. TikTok requires realistic AI-generated content to be labelled, and its Ads Manager offers ad disclaimers, including one for AI-generated content. Policies differ per platform and change, so read the live policy where you buy.

The EU has statutory transparency rules. The European Commission's guidance on transparency obligations under Article 50 of the AI Act covers AI-generated or manipulated image, audio and video content, and states that the article applies from 2 August 2026. If you advertise in Europe, budget the labelling into the creative, not into a legal review afterwards.

And the rule that predates all of them: the face must be yours to use. A real person's likeness needs written consent, and public figures are off limits.

How Sosie fits

Sosie sits in one of those four boxes and not the others. It does not make image ads, it does not write your copy, and it does not invent scenes from a prompt. It does performance capture: you record up to forty seconds on your phone, saying your own script, and Sosie puts that performance on a presenter you choose from the catalogue, describe in words, or upload as an image you hold the rights to. Your voice, timing, expressions and gestures are kept. By default the presenter appears in the room you recorded in, with the same walls and light; a presenter backdrop is the alternative.

What comes back is a vertical 1080p MP4, landscape if you filmed landscape, with no watermark and cleared for commercial use including paid placements. One credit covers a video of up to twenty seconds and two cover up to forty. Credits are sold as monthly subscriptions or as one-time packs that never expire, from €39 for five videos, and no plan exists without payment. A render that fails is refunded to your balance. Because your own audio drives the render, the presenter speaks the language you spoke.

The terms require consent for any real person you upload and prohibit public figures and deceptive use, and the product's guidance is the presenter-and-demo register described above. If you want the broader picture first, the guide to what UGC is covers where this style of ad came from, and the guide to AI actors goes deeper on the presenter itself.