Seedance 2.5
A cinematic video model. Long scenes and natural camera movement.
Text to videofrom 14 credits / s
Video generation
Generate AI video from a prompt, a still, or reference images
Vellria generates AI video on prepaid credits with nothing to renew: pick a video model from the catalog, write the shot, and get a finished clip back for the amount that model's page stated before you started. A clip can begin from a written description alone, from a still that becomes its opening frame, or from reference images that fix what the subject looks like, and those are three different models in the catalog rather than three settings on one.
A clip renders as a single take
A video model does not build a timeline. It renders the whole clip in one pass, and there is no point mid-render where you can step in, cut a beat, or change your mind about the camera. What you asked for arrives complete or it does not arrive. That makes the prompt the edit: the action you describe is the action you get, first frame to last.
So write the whole shot before you run it, not the first half of it. Say where the subject starts, what changes, and where it ends up. Then iterate by re-running: change a verb, a light, or a camera word, and compare the takes. Steering happens between runs, which is why it pays to settle the framing and the movement on a short draft before you commit to a long clip at the top of a model's range. Billing and the mechanics of starting a run are the same for video as for stills, and the image generation page sets them out.
Models
Seedance 2.5 Image
Takes the image you provide as the first frame and animates it into a long scene.
Image to videofrom 14 credits / s
Seedance 2.5 Reference
Builds a scene from a set of reference images, keeping the subjects consistent.
Reference to videofrom 14 credits / s
MiniMax H3
Fast, economical video generation. Straight from text to scene.
Text to video6 credits / s
MiniMax H3 Image
Takes the image you provide as the first frame and animates it.
Image to video6 credits / s
MiniMax H3 Reference
Generates video that keeps character consistency across several references.
Reference to video6 credits / s
HappyHorse 1.1
A video model that reaches 1080p. Nine aspect ratios, 3–15 seconds.
Text to videofrom 7 credits / s
HappyHorse 1.1 Image
Takes your image as the first frame and animates it up to 1080p.
Image to videofrom 7 credits / s
HappyHorse 1.1 Reference
Takes up to nine references and generates video that preserves subject and scene.
Reference to videofrom 7 credits / s
Wan 3.0
The cheapest video in the catalog. A single scene up to 30 seconds, with audio.
Text to videofrom 5 credits / s
Wan 3.0 Image
Takes your image as the first frame and animates it for up to 30 seconds.
Image to videofrom 5 credits / s
Wan 3.0 Reference
Takes up to ten references and generates video that preserves character and place.
Reference to videofrom 5 credits / s
Wan 3.0 Pro
The large version of Wan 3.0. Higher detail and steadier motion.
Text to videofrom 7 credits / s
Wan 3.0 Pro Image
Starts from the first frame and continues with Pro's detail and motion quality.
Image to videofrom 7 credits / s
Wan 3.0 Pro Reference
Combines ten references at Pro quality; identity and place are preserved.
Reference to videofrom 7 credits / s
Motion has to be written down, not implied
A still prompt describes a frame. A video prompt has to describe a frame and then what happens to it. The subject moves and the camera moves, and a model handed nothing but scenery will decide both on your behalf. Walks toward the door and turns away from the door are different clips with identical set dressing, and the whole difference lives in the verb you wrote.
Camera language is the part most people leave out. A slow dolly, a locked-off frame, a handheld drift and a push-in read as different shots, and omitting the camera does not give you a static one, it gives you an arbitrary one. Pacing works the same way. Say whether the action is a single beat or a sequence of them, because the model has to fit whatever you described into the length you asked for, and an action written for twice that length arrives rushed.
The three routes answer different questions
Working from a description alone is for a shot that does not exist yet: you are asking for a composition as well as for movement, and everything in the frame has to be in the text. Working from a still is for a shot you have already framed. The composition is settled before the run begins, so the whole render goes into motion rather than into deciding what the scene looks like.
Working from references sits between the two. The pictures are evidence of who should appear and where, but none of them becomes a frame, so composing the shot is still the model's job and the staging is still yours to write. What you attach, in other words, is a statement about which decisions you have already made. Attaching more of them is not a quality upgrade; it is a narrower question.
The frame shape is decided before the run, not after
Some video models take an aspect ratio as part of the request, and on some of those it is required rather than optional. Send a ratio to a model that does not offer one and the request is refused rather than silently accepted, which is the behaviour you want: a clip that came back in a shape you did not choose, with no error to explain it, is worse than a rejected request.
Where a still becomes the opening frame, that picture fixes the shape too, and there is no aspect ratio left to pick afterwards. So crop before you upload. Cropping the finished clip instead means throwing away pixels the model spent your credits rendering, and re-framing a composition that was built for the wider box.
Continuity across shots is the hard part
A clip on its own is the easy case. A run of clips that look like the same production is not, because nothing carries between runs by itself. Each take is generated from scratch, so a character can return with a different face, a room with different furniture, a key light from a different side, even when the wording barely moved. That is the work, not a defect in any particular model.
The catalog gives you handles on it. Reference images tell a model what the subject looks like instead of asking it to invent an equivalent every time, which is what holds a face steady from shot to shot. A still used as an opening frame does the same job for one specific moment: make the frame on an image model, keep it, then animate it. And keep your wording stable between takes, since a rewritten description is a new negotiation over everything you did not intend to change.
Frequently asked questions
How is video billed?
By the run, out of a prepaid balance, with nothing renewing between projects. What a clip costs scales with the length and the resolution you asked for, and the model page names the amount before you start rather than after. A run that ends in a provider error returns those credits to the balance on its own.
Can I trim or extend a clip after it renders?
Not inside the generator. A take runs start to finish once it begins, and there is no mid-render checkpoint to cut at or continue from. You change the result by changing the prompt and running again, or by taking the downloaded file into your own editor. Plan the clip as one complete take before you commit to it.
Can I animate a picture I already have?
Yes, on the models that take a still and use it as the opening frame. Crop it first, because that picture also fixes the shape of the clip. The models that take references work differently: they read your pictures for what the subject looks like without turning any of them into a frame, so the shot still comes out of your prompt.
Why does my character look different in every clip?
Because nothing carries between runs. Each take is built from scratch, so a description alone is re-negotiated every time and lands somewhere slightly different. Supply the character instead of describing it: feed the same reference images, or the same opening frame, into every take in the sequence, and keep the wording steady around them.
How do I decide how long a clip should be?
Write the action first, then match the length to it. The model has to fit whatever you described into the duration you asked for, so a single beat stretched too long drifts and a sequence squeezed too short arrives hurried. Each model page lists the durations that model accepts, and the range is not the same across the catalog. The limits on what a clip may depict are the same as everywhere else here and are set out on the content rules page.