Guide
Image-to-video or text-to-video?
The difference is not quality. It is who decides the first frame — and that one choice determines how much of the result you can control.
The actual difference
Who owns frame one
Text-to-video takes a description and produces a clip. It invents everything: who is in it, what they look like, where they are, how the camera sits. You get motion and a scene from a sentence, which is remarkable, and you get whatever it decided, which is the problem.
Image-to-video takes a picture you supply and animates it. The first frame is yours — exactly the person, exactly the framing, exactly the light. The model’s job is narrowed to working out what happens next.
Everything else follows from that. The question is never which is better; it is whether you have, or can make, the right first frame.
Why it matters for characters
Identity work starts from an image, nearly always
If a specific person has to appear in a clip, text-to-video is working against you from the first frame. It is being asked to conjure a face from a description, which is the problem that has no good solution in words, and then to hold that invented face for several seconds of motion.
Image-to-video sidesteps it. You establish the character in a still — where it is cheap to iterate, cheap to reject, and where you can check the face properly — and only then spend the much more expensive video generation on animating something already correct.
That is also simply a better use of money. A still that is wrong costs seconds to replace. A clip that is wrong because the face was wrong in frame one costs the whole generation, and you find out at the end.
When text-to-video is the right call
It has real uses, and they are not identity
Nobody specific in frame. Establishing shots, landscapes, weather, abstract motion, crowds at a distance. If no face has to be the same face as anything else, the constraint that makes text-to-video hard does not apply.
Exploring, not executing. When you do not yet know what you want, describing it and seeing what comes back is faster than making a still first. Use it to find the idea, then rebuild the shot properly once you know.
Motion that is hard to stage. Some camera moves and some physical events are easier to ask for in words than to imply from a single frame, because a still gives the model no information about what was about to happen.
The trade you are making
A first frame constrains more than the first frame
Image-to-video is not free of cost. The supplied frame anchors the whole clip, which is the point, and it also limits it. A tightly composed still gives the model nowhere to move the camera to, so clips often feel locked off. A still with a subject mid-gesture can produce motion that reads as continuing; a static posed portrait often produces very little motion at all, because nothing in the frame implies momentum.
Choosing the starting frame is therefore a directing decision rather than a technical one. Pick the moment just before the action rather than the most attractive still, leave room in the frame on the side the subject will move toward, and avoid a pose that looks complete — a completed gesture has nowhere to go.
In a sequence
This is also how shots get joined
Once a pipeline is doing image-to-video, the last frame of one clip is available as the first frame of the next, which is exactly the frame chaining the video consistency guide describes. Continuity across a cut stops being something you hope for and becomes something the structure gives you.
That is the real reason longer work is built this way rather than generated in one go. Text-to-video gives you a clip; image-to-video gives you a clip that can be the start of the next one. Unlimited Studio shoots on the second pattern for that reason, with the character reference applied to each shot so the chain does not drift away from the person it started with.
Keep reading
The rest of the series
- Why AI video looks floatyThe weightless, slow-motion tell: where it comes from and which parts you can actually prompt away.
- Why hands and faces come out wrongTwo different failures with two different causes - and why asking for perfect hands does nothing.
- Why the same prompt gives a different pictureA seed is one of five variables. Why fixing it is necessary but not sufficient for reproducing an image.
Where this comes from
Unlimited Studio is built around this one problem
Everything above is what the product does: a character is locked from eight reference angles, each generation conditions on whichever angle its prompt implies, a score checks the result against the reference, and a film is shot one frame-chained shot at a time. What it will and will not generate is set out in the content policy.