Text-to-video and portrait-to-video are both called "AI video", which is why people are confused about why one of them never looks like them. The difference isn't quality. It's what the model is given.
Two different starting points
Text-to-video starts with a description. The model invents everything: people, faces, room, light. If you write "a zombie couple in a candlelit chapel," you get a couple — not yours. There is no mechanism by which your faces enter, because nothing about your faces was supplied.
Portrait-to-video starts with identity. Two photographs, one per person, plus a scene description. The model has to render what it was given rather than invent it.
Everything downstream follows from that.
Why text-to-video can't be fixed with better prompts
People try. They describe a face in detail — jawline, eye colour, hair, freckles. It doesn't converge, for two reasons:
- Description is lossy. A face is hundreds of parameters, and language carries maybe a dozen.
- Consistency isn't enforced. Two clips from the same prompt produce two different people. Within a clip, the face can change between frames.
A prompt can specify a type. It cannot specify a person.
What portrait-to-video still gets wrong
To be fair to the comparison: feeding two photos isn't a guarantee either. Clothing shifts, jewellery drops out, a hairline moves. The difference is the class of error — the model starts from your face and degrades it, rather than never having it.
Degradation you can plan around: simpler wardrobe, even light, ten seconds, locked scene. Absence you cannot.
Where text-to-video genuinely wins
- Scenes without specific people. Establishing shots, landscapes, a horde in a street. No identity to preserve, so nothing to lose.
- Concept work. Testing an idea before committing to anything.
- Weirdness. If you want something no reference image implies, text is the only entry point.
- B-roll. Cutaways that make a sequence feel bigger.
The hybrid that professionals actually use
Text-to-video for the world, portrait-to-video for the people. Generate the environment as a text clip, generate the couple as a portrait clip, cut them together.
This works because each tool is being used inside its competence: the general model does what it's good at (inventing places), and the portrait route does what only it can (keeping two faces).
What about "image-to-video" as a term?
Mostly it means one image as the first frame, which is a middle case: it anchors composition strongly and identity somewhat, but a single frame tells the model nothing about how that person moves or what they look like from another angle.
Two portraits plus a scene is a narrower, more opinionated input. That narrowness is why it works for this job and why it's useless for a general film.
Cost comparison
Text-to-video arrives inside subscriptions with credit systems, which makes the true cost of a usable clip hard to know in advance. A per-video product fixes the price at $3.99 and the output at 10 seconds, 16:9, 720p.
If you generate daily, subscriptions are correct. If you want one video of two specific people, per-video is correct and the subscription is a rounding error you'll forget to cancel.
The one-line rule
If the sentence you'd write starts with "a couple," text-to-video is fine. If it starts with "us," you need portraits.
Try it at the generator; price at pricing.
Frequently asked questions
Can text-to-video make a video of specific people?
Not reliably. It can be prompted toward a description, and some products accept an image reference, but identity is not guaranteed by the text path.
Is portrait-to-video higher quality?
Not necessarily in absolute terms — it's narrower. It's better at the one thing it's for: two recognisable faces in one scene.
Can I supply more than two photos?
This product takes one portrait per person. More inputs don't improve identity here, and a clear front-on face matters more than volume.
What's delivered, exactly?
A 10-second, 16:9, 720p MP4, downloadable from a private task link, with an email notification when it's ready.