Seedance 2.0 Explained: How a Still Photo Learns to Move
Ozilai TeamA plain-language guide to Seedance 2.0: how text, a first-frame photo, or reference material becomes a moving clip, what it costs, and what it cannot do.
Seedance 2.0 is the video generation model that powers Ozilai.video. You give it one of three things — a text prompt, a first-frame image, or reference material — and it returns a short clip in which things move. The input you choose is the decision that shapes everything downstream, because it sets how much of the scene the model invents and how much of it is anchored to something you supplied. It is not an image model that repaints a picture; it renders motion. And there is a short list of things it genuinely cannot do, which is worth knowing before you spend a credit.
Générez le vôtre ici
Choisissez une photo claire et bien éclairée de la personne à qui appartient le but.
Téléchargez une photo pour continuer.
En générant, vous confirmez que vous avez la permission d'utiliser cette photo.
Votre clip apparaît ici
Un rendu avec les paramètres par défaut coûte 30 crédits.
The Three Kinds of Input, and What Each One Returns
The fastest way to understand any generation model is to look at what you hand it, what it does with that, and what comes back. Seedance 2.0 accepts three broad categories of input, and they are not interchangeable — each one trades control for freedom in a different place.
A text prompt is the loosest input. You describe a scene and an action, and the model builds everything: the people, the setting, the light, and the movement. Nothing anchors identity or framing, so you get a clip that captures the mood and the action of your description while the specific people in it are invented. This is the right input when you want a scene rather than a specific person.
A first-frame image is the tightest anchor. You supply a picture, the model treats it as the opening frame of the clip, and it generates the motion that follows. The person and the composition start exactly where your photo put them, which is why image-to-video is the route that keeps a real face recognisable. The trade-off is drift: the further the clip travels from its first frame, the more the faces and background wander away from your original.
Reference material sits between the two. Instead of being the literal first frame, a reference steers how the result looks or how it moves. You use it when you want a clip to echo a particular visual texture or a particular style of motion without being a frame-for-frame copy of anything.
| Input you give | What the model does with it | What to expect back | | ----------------------------------------- | ----------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------- | | A text prompt only | Builds the whole scene from your words — subject, environment, light, and camera | Motion that matches your description, with invented people. Ideal for scenes, useless for a specific face | | A first-frame image | Treats your picture as frame one and extends it forward into movement | A clip that begins on your photo and moves away from it; the longer the clip, the more the detail drifts | | Reference material | Uses the reference to steer look or motion rather than as the literal opening frame | A result that echoes the reference's texture or movement while keeping its own composition | | A first-frame image plus a written prompt | Anchors identity and framing with the image, then directs the action with the words | The most controllable output: the person you supplied, performing the move you described, in the setting you asked for |
Why the first frame carries the most weight
When you hand the model a still image, you have answered the hardest questions in advance: what does this person look like, and where are they standing in the frame. Everything the model generates afterwards has to stay consistent with that answer. That is a constraint, and constraints are what make results controllable. It is also why a tight, well-lit, single-subject photo outperforms a wide snapshot — the stronger your frame one, the more the model has to hold onto as it invents the rest.
What reference material is actually for
Reference input is often misunderstood as "a picture of the thing I want". In practice it is closer to "a note about how I want it to feel". You reach for it to borrow a texture, a colour palette, or a genre of camera movement — the grain of a broadcast, the look of a certain era of television. Treat it as direction, not as a source of truth, and you will use it well.
Why a Motion Model Is Not an Image Model
This distinction matters more than any prompt trick, and it explains most of the strange failures people see. An image model takes a picture and changes how it looks: it restyles it, relights it, upscales it, or reimagines it in a different medium. The input and the output are the same shape — one frame in, one frame out — and every part of the output corresponds to a part of the input. Nothing has to be consistent with anything, because there is nothing else. That is why an image model can turn a portrait into a painting with almost no risk of the process going wrong.
A video model has a different job. It has to produce a sequence in which the world stays coherent from the first frame to the last. The same face has to survive twelve seconds of movement; arms have to bend in directions arms bend; a ball has to travel along a path rather than teleport; the light has to keep falling from the same side. The model is not deciding what a picture looks like — it is deciding what happens next, over and over, and then holding all of those decisions together.
Consistency across time is the hard part
Every visible artefact in AI video traces back to that consistency problem. A face that melts two seconds in is a model that lost its grip on identity somewhere in the middle of the clip. A figure with an extra limb is a model that resolved two overlapping guesses about where a body was. Shimmering light is a model that made a slightly different decision about the sun on consecutive frames. None of these are mistakes in the ordinary sense; they are the price of the model having to invent a plausible future and keep it stable.
The practical consequence is worth internalising: a video model is choosing among possibilities, not copying a source. The clearer your anchor and the fewer simultaneous things you ask for, the fewer choices the model has to make, and the more consistent the result. That is the entire logic behind writing short, ordered prompts and using tight photos.
What Resolution and Duration Actually Buy You
The two numbers that shape both the cost and the believability of a clip are its frame size and its length. They work in the same direction: more detail and more time both cost more, and both make the result more demanding to hold together.
On a phone screen, a 720p clip already reads as convincingly photographic — a face is clear, a shirt moves, a ball travels. The case for 1080p shows up when the frame is large or when you intend to crop into it. Cropping a 1080p clip to a vertical short-form ratio discards a lot of pixels; starting from a higher-resolution frame means the subject survives that crop with detail left over. If you are posting vertical video, the larger frame is doing real work rather than just looking tidier in a file browser.
| Output | How long | Credits | | ---------------------- | -------- | ------- | | 480p — lighter option | 12s | 5 | | 720p — standard video | 12s | 10 | | 1080p — studio default | 12s | 30 |
Duration works the way you would expect: the price scales with the length of the clip and is rounded up to a whole credit, so a short clip never costs a fractional amount. Aspect ratio, by contrast, is free — rendering the same clip as vertical 9:16 or widescreen 16:9 does not change the credit cost. That means the format decision is yours to make on the strength of where the clip is going, not on price. The credits explainer breaks the arithmetic down plan by plan.
What Seedance 2.0 Cannot Do
Honesty about limits is more useful than a longer feature list, because most disappointment comes from expecting something the model was never going to deliver.
- It will not reproduce a specific real broadcast. If you ask for a famous match moment exactly as it aired, there is no footage in the model to replay. It will generate something in the spirit of that kind of moment — the same shape of action, the same broadcast feel — but not a copy of the original. This is precisely why the photo-to-video trend works the way it does: you rebuild the choreography in the prompt rather than expecting the model to remember a match.
- It cannot invent a person's likeness from nothing. A written description produces a generic person who fits the description. A specific face has to arrive as an image; that is what it means to supply an anchor. If you want a recognisable person in the clip, the photo is not optional.
- It cannot guarantee a particular camera move. You can ask for a long lens tracking left to right and you will usually get something that respects that request. But a camera instruction steers the result rather than commanding it. When an exact move is essential — the operator not crossing the line, the goalkeeper staying in shot — plan on more than one attempt and treat the first render as information.
You can also add a fourth expectation to that list, though it is a limitation of scale rather than of capability: the model will not build a minute of continuous, coherent action out of one still. Short clips are where it is strong, and the honest way to make something longer is to render several short pieces and join them, rather than to stretch a single generation past the point where it can hold its world together.
How to Get a Usable Clip Out of One Generation
The workflow below is the shortest path from an input to a result you would actually post, and it works whether your input is a photo, a prompt, or both.
- Decide what you are making: a scene or a person. If specific identity matters, image-to-video is your route and a photo is mandatory. If you only need atmosphere or action, a text prompt is enough.
- Strengthen the anchor before you touch the prompt. One subject, head-and-shoulders framing, even light, sharp focus. This single step changes results more than any prompt rewrite.
- Write the action as ordered beats. One clause per beat, in the order they happen. A striker who receives the ball, then chips the goalkeeper, then finishes is three beats.
- Add exactly one camera instruction. Two camera instructions fight each other and produce mush; one steers cleanly.
- Pick frame size and length, then check what it costs. The standard 720p, 12-second render is 10 credits; the studio default of 1080p costs 30 at the same length.
- Generate, compare against your beats, and change one variable. Keep the photo if you rewrite the prompt; keep the prompt if you swap the photo. Single-variable iteration is what turns a budget into progress.
If your goal is the football trend specifically, the step-by-step Ozil AI video guide and the prompt guide for that goal apply this workflow beat by beat. If you want to see the same machinery pointed at other footage, photo-to-football-highlight shows how far one still can be pushed.
FAQ
Is Seedance 2.0 just a photo editor with extra steps?
No. A photo editor alters the picture you give it and hands the picture back. Seedance 2.0 takes a picture and produces a sequence of new frames in which the world moves. The output is a clip, not a modified still, and the interesting decisions it makes are about motion rather than about appearance.
Can Seedance 2.0 recreate a real football broadcast exactly?
It cannot replay footage that exists. What it can do is generate a clip in the visual language of a broadcast — the framing, the pace, the choreography — with the people you supply. That is why recreating a famous goal means describing the sequence in the prompt rather than asking the model to remember it.
Do I need to write a long, detailed prompt?
Long is not the same as detailed. Five ordered beats and one camera instruction will outperform a paragraph of adjectives, because every extra simultaneous demand is another thing the model has to keep consistent across the clip. Write the action in order, keep the subject count low, and stop.
What happens if the generation fails?
A failed render is refunded automatically to the credits it was drawn from, so you are not charged for output you never received.
Which resolution should I try first?
Start at the standard 720p render while you are learning what the model does with your photo — it costs 10 credits and tells you almost everything you need to know about the crop and the prompt. Move to the 1080p studio default once the action is right and you want the extra detail, especially if the clip will be cropped for vertical video. Details on plans and pricing are on the pricing page.
Is Seedance 2.0 the only model that works this way?
The same principles apply to any modern video generation model: the input anchor determines control, time consistency is the hard problem, and the limits are about replaying reality rather than about imagination. Learning the model's categories once means you can read your way through any of the alternatives, which is what our comparison of Ozil AI video alternatives does.