Video
WAN 3.0
Fast iteration for short-form content.
Available on Ofrai — chat, images, and video in one place with shared credits.
Example prompts
Choose an example to load its prompt and recommended settings into the WAN 3.0 creator above.
How it works
Create with WAN 3.0 in three steps
- 01
Choose your inputs
Keep WAN 3.0 selected, then add a prompt and any supported reference assets.
- 02
Set the output
Choose the aspect ratio, resolution, and available Video options before generating.
- 03
Generate and refine
Review the result, adjust the prompt or settings, and generate another version from the same place.
Wan 3.0 is the all-purpose surface of this group: a brief can arrive as words alone, as a single opening frame, as an opening and a closing frame, or as a bundle of image, video and audio references submitted together — and the endpoint accepts all four shapes rather than routing between them. That breadth is the reason to pick it when the deliverable is not yet decided.
The output window runs from 2 to 30 seconds with a smart duration of -1 available, resolution climbs 480p, 720p and 1080p with 1080p as the documented default, and audio generation is on unless you switch it off.
Two things about this endpoint are less obvious than the capability list suggests: the prompt field is written to describe sound as well as picture, and the reference documents no aspect ratio parameter at all, which means the frame of a Wan 3.0 job is decided by your assets rather than by a field you can set.
Four shapes of brief, one endpoint
The interface description lists four ways a job can enter, and they are not variations on one another — they constrain the deliverable in different ways. Text-to-video is the origin case, where the model invents subject, setting and motion. First-frame control hands the composition over to a still you art-directed and lets the model spend its effort on what happens after it.
First-and-last-frame control pins both ends of the take, which is the only documented way here to guarantee a specific closing image, and it is the mode that decides transitions: if the last frame is the product on the shelf, everything between the two frames is written to arrive there. The fourth shape is joint generation from reference image, reference video and reference audio submitted together, which is the mode where several kinds of material act on one shot at once rather than one of them anchoring it.
Reference to file or public web page content sits in the same list. Read the list as a triage of briefs: four frames means a defined arrival, a reference bundle means a defined look, a still means a defined opening, and words alone means a defined subject.
The prompt carries the sound, not just the picture
Read the prompt parameter description literally and it asks for four things: the content of the video, the subject's action, the camera, and the sound. That last item is the one most callers leave out, and it matters because sound generation on this endpoint is a separate boolean that defaults to on. You therefore make the sound decision twice, in two different places, and the two can disagree. The boolean decides whether a track is rendered at all; the sentence in the prompt decides what is on it.
A request that leaves the sound unspecified with the boolean on gets generated audio that no one wrote down, which is the least controllable outcome available. Say what the room sounds like in the same breath as what the camera does, then set the boolean to match your edit — on for a clip that carries its own ambience, off for a plate that will sit under dialogue.
The prompt is also optional, but not optional alone: the documented precondition is that the prompt and the media assets provide at least one between them, so an asset-led job is valid and a request carrying neither is not.
One field that serves a two-second bumper and a thirty-second scene
Duration accepts anything from 2 to 30 seconds, defaults to 5, and additionally accepts -1, which hands the length decision to the model so it can read the prompt and the supplied material and choose. The spread is unusual enough to be worth planning around, because the same parameter has to cover both ends of an edit.
At the short end, two-second outputs are real requests rather than a rounding artefact, and they are the right length for a title sting, a beat between two cuts, or a loop plate — none of which want the last three seconds trimmed off.
At the long end, thirty seconds is long enough for one generation to carry a short scene rather than a single shot, which is exactly where a sequence is most tempting and most fragile: a single continuous take that holds subject, background and camera for half a minute is beyond what one take reliably manages, so plan the cut points before the request rather than after the render.
The two-second and thirty-second requests share one field, so decide which job you are on first; -1 is a third option rather than a fourth, and it is the one to reserve for material whose length you genuinely cannot predict.
Building the reference video array before you submit
Reference video goes into video_list, and the structure of an entry is the first thing to get right: each item is an object carrying a video_url, so a bare URL string in the array is a malformed request rather than a shorthand. Two ceilings apply and they are different in kind. Five is a count limit on the array; fifteen seconds is an aggregate across everything in it.
Five five-second clips fill the budget exactly, and a single twenty-second clip does not fit at all no matter how few items the array holds. Check the aggregate before you submit rather than discovering it as a rejection, because trimming a reference set is a decision about which motion matters and re-deriving that from a failed call wastes the round trip.
The reference budget is also independent of the output duration, which is the second thing people conflate: fifteen seconds of reference material and a thirty-second output are two separate limits, and asking the model to stretch a full reference list out to thirty seconds is a demand on its judgement rather than on its compute. Every asset in this group has to be a reachable URL before the call, which makes hosting a step in the pipeline rather than an afterthought.
What this endpoint does not document
Two gaps are worth stating plainly, because building against an assumption the reference does not make is how integrations break. First, the documented parameter set for Wan 3.0 covers duration, resolution, prompt, video_list and audio — there is no aspect ratio parameter among them.
The frame of a Wan 3.0 render is therefore derived from the material you supply rather than chosen up front, and a job that requires a specific geometry has to get it from a source frame, a reference clip or a source video whose own proportions are correct. Do not send a ratio field out of habit and assume it was honoured.
Second, the capability list includes reference to file or public web page content, but it is stated in one line alongside the others without a parameter of its own and without a payload shape, while video_list above it does have one. Verify the exact request structure for that capability against the current reference before building on it, and treat the capability as promising rather than as specified until you have seen the shape. Both gaps are cheap to design around and expensive to discover in production.
When Wan 3.0 is the wrong instrument
Breadth is not free, and the three costs of this endpoint are specific. The first is that you cannot ask for a frame shape, so any job where geometry is the brief rather than a by-product belongs somewhere else. The second is that a 1080p default means a draft is a top-tier draft unless you override it; set 480p while the prompt and the framing are still moving, then move up once.
The third is that the reference budgets are tight — five clips, fifteen seconds total — which suits a look you can point at rather than an archive you want the model to study. On the neighbouring models, Wan2.6 is the opposite shape: it routes by payload, offers only 5, 10 or 15 seconds, restricts resolution to 720p and 1080p, and takes a single image URL. Seedance 2 reaches 4k on its base id. Seedance 2.5 adds editing and extension tasks, which Wan 3.0's reference does not describe.
If your brief needs to change an existing clip rather than make a new one, that is a Seedance 2.5 job; if it needs a 4k master, it is a Seedance 2 job; if it needs 2-to-30-second flexibility with first-and-last-frame control and reference video, it is this one.
FAQ
- What does passing -1 for duration actually do?
- It is a smart-duration request: the model reads the prompt and the supplied material and chooses the output length itself. It is a value you send, not an absent field — the documented default when duration is omitted is 5 seconds, so -1 has to be sent explicitly to get the model's judgement instead of the default five.
- Is the prompt required on Wan 3.0?
- Not on its own. The documented precondition is that the prompt and the media assets provide at least one between them, so a request carrying only an image or only reference video is valid, and so is one carrying only a prompt. What fails is a request carrying none of them. Validate that union rather than requiring the prompt unconditionally, or you will reject asset-led jobs that the endpoint would have accepted.
- How many reference videos are allowed, and what is the payload of one entry?
- video_list takes at most five entries whose combined duration is fifteen seconds or less. Each entry is an object with a video_url field pointing at the clip, not a bare URL string. The five-entry count and the fifteen-second aggregate are separate limits, and both are independent of the 2-to-30-second output duration.
- Can I set the aspect ratio?
- There is no aspect ratio parameter in the documented parameter set for this endpoint, which covers duration, resolution, prompt, video_list and audio. The frame is derived from the material you supply, so shots that need a specific geometry have to arrive with it — through a source frame, a reference clip or a source video with the right proportions.
FAQ
- Can I use WAN 3.0 directly on this page?
- Yes. The creator above is the working Ofrai studio with WAN 3.0 selected by default.
- Where are my finished results saved?
- Signed-in generations are kept in your Ofrai library so you can download or revisit them.
- Which prompts work best with WAN 3.0?
- Describe the subject, setting, visual style, lighting, and camera or motion clearly. The examples above provide practical starting points.
- Can I edit the example prompts before generating?
- Yes. Load any example, then change its wording, aspect ratio, references, or other supported settings in the creator.
Related models
Compare other Video models that can handle similar creative work to WAN 3.0.
Related guides
Tools
Commercial use depends on the selected model and model terms.





