How to Use the Wan 3.0 Video Generator: Inputs, Prompts and Budgets
The 30-second Wan 3.0 take starts long before the prompt: it starts with choosing the right input. Alibaba's model does not only read text — it reads a script you upload, a web page you link, and up to ten images, five videos and five audio clips you attach. This guide is the operator's companion to our Wan 3.0 30-second guide (which covers the specs and how to plan a long take): here we go through each input, the prompt format that holds 30 seconds together, what a generation costs at every resolution, and the mistakes that waste money. Last checked: 2026-10-10.
The five inputs, and what each is for
| Input | Limit | What it is for |
|---|---|---|
| Document | 1 file, up to 100 MB and 50 pages | The source of truth for content: a script, a brief, a storyboard or a slide deck, passed in directly instead of being pasted into the prompt. |
| Web link | 1 link | Context that lives on a page — a product listing, an article, brand documentation — when you do not have a file. |
| Images | Up to 10 | Identity anchors: the character, the product, the location, the style frame. They are what keeps the last second looking like the first. |
| Videos | Up to 5 clips, 15 s total | Motion and continuity: how the subject moves, how the camera behaved in a previous take, a clip to extend. |
| Audio | Up to 5 clips, 15 s total | Sound references: a voice to keep consistent, a jingle, ambience to match. |
You pass one document or one link, not both, and the media limits are hard caps the API enforces. A practical split: the document or link carries what happens, the images carry who it happens to, and the prompt carries how it is shot. The prompt itself can run to 20,000 characters, so a beat sheet that does not fit the document still fits the prompt.
Prompt structure for a long single take
Alibaba's own guide formats long prompts as timestamped segments, each with a camera line and a scene line, under global style and audio lines. The format matters more than the length:
Style: warm documentary realism, golden-hour light, 16:9.
Audio: quiet street ambience, soft acoustic guitar under the dialogue.
(0:00 - 0:08) Camera: wide shot, slow push-in. Scene: a cyclist rides into an empty market square.
(0:08 - 0:20) Camera: continues to a medium shot, handheld follow. Scene: she stops at a flower stall. Dialogue (vendor): "You're early today."
(0:20 - 0:30) Camera: slow orbit to a close-up. Scene: she smiles and picks up a bunch of sunflowers.- Global lines first. Style, lighting and the audio bed apply to the whole clip; state them once at the top, not per segment.
- One action per segment. Segments that stack two actions are the ones the model averages into mush.
- Name the camera transition. In a single take there are no cuts — write how the camera travels from one beat into the next ("continues to a medium shot"), not just where it ends up.
- Budget the dialogue. Natural speech runs two to three words per second; leave air between lines instead of filling all 30 seconds.
- Choose the duration deliberately. The default is 5 seconds; -1 asks the model for a "smart duration" it picks itself. When the clip has to land on a music cue or a 15-second ad slot, set the seconds explicitly.
More worked examples of this format are in our Wan 3.0 prompt collection, and the Wan 3.0 overview summarises the model's capabilities.
What a generation costs: worked examples
Wan 3.0 bills per second of output, and the rate depends on resolution and tier. The standard model (wan3.0-video) lists at $0.05 / $0.10 / $0.20 per second at 480p / 720p / 1080p on the international price list; wan3.0-video-prime, the faster tier with the same capabilities, lists at $0.068 / $0.14 / $0.28. Turning the native audio off does not change the price.
| Clip | Standard | Prime |
|---|---|---|
| 5 s at 720p | $0.50 | $0.70 |
| 10 s at 480p | $0.50 | $0.68 |
| 10 s at 720p | $1.00 | $1.40 |
| 30 s at 480p | $1.50 | $2.04 |
| 30 s at 720p | $3.00 | $4.20 |
| 30 s at 1080p | $6.00 | $8.40 |
Two habits follow from the arithmetic. First, draft cheap and finish sharp: a 30-second draft at 480p costs $1.50 against $6.00 at 1080p, so settle timing and blocking before the final render. Second, pay for Prime only when you are waiting on the render — in overnight batches the standard tier produces the same clip for less. Wan 3.0 can also extend an existing clip forwards or backwards, as long as the input and output together stay within 30 seconds, which is usually cheaper than regenerating.
Common failure modes
- The document is rejected. Over 100 MB or over 50 pages and the request fails before generating. Split the deck or link the page instead.
- Reference limits are hard. The eleventh image, a sixth video clip, or reference video that runs past the 15-second total is not "risky" — it is refused. Curate down to the anchors that matter.
- Identity drift on long takes. Thirty seconds is a long time for a face or a label to stay put. Pass the character or product as reference images and check the last third of the clip, not the first.
- Smart-duration surprises. With duration -1 the model picks the length, which is fine for exploration and wrong for a 15-second ad slot. Set the seconds when the output has a deadline.
- Extending past the cap. Extension counts input plus output: a 20-second input cannot be extended by 15 more seconds. Budget the extension inside 30 seconds.
- Expecting weights. Wan 3.0 is API-only and publishes no weights. If you need an open model you can host, the family's last open flagship is Wan 2.2 under Apache-2.0.
What you can use on Imgo today
Wan 3.0 is not integrated into Imgo and has no Imgo credit price — this page is an external reference. What is live on Imgo is the Seedance 2 family: clips of 5 to 15 seconds, 480p up to 4K on Seedance 2.0 (480p and 720p on the Fast and Mini tiers), up to 4 reference images, 3 reference videos and 3 audio tracks, and native audio generated with the picture. Two constraints worth knowing: reference images cannot be combined with start or end frames, and an end frame needs a start frame. Billing is per second at the model's rate — Seedance 2.0 follows Volcengine's official rate, and Seedance 2 Fast costs about 55% of it — and the generator quotes the exact credit price before you submit, every run.
To see everything that can be generated right now, check the live model list; for the other options in Wan 3.0's class, our Wan 3.0 alternatives page and the Wan 3.0 vs Seedance 2.5 comparison set out the field.
Sources
Share
Related Posts
No image
Kling 4.0 Flash opened in early access on September 28 for a limited group — reported as Ultra yearly subscribers — with 3–20 second 720p clips. Here is how to get in, what it costs, how to write the 8,000-token prompts and keyframe plans the full model will accept, and what to generate meanwhile.
No image
Wan 3.0 accepts a document, a web link, images, videos and audio as inputs. This operator's guide covers what each input is for, how to structure a 30-second-take prompt, per-second budget examples at 480p to 1080p on both tiers, and what commonly goes wrong.
No image
Kling 4.0 Flash has been in early access since September 28, 2026; the full model and its API are pending, with an October 2026 target. Here is what each tier does, which headline specs are only announced, and what to use in the meantime.