Wan 3.0 Prompt Guide: Structure, Examples and Real Clips

Write a Wan 3.0 prompt in Alibaba's order: an overview line, any references, numbered shots with start and end times, dialogue in quotes, sound and music, style, then a short list of what must not appear. We ran five prompts at 720P on October 27, 2026: the timed cuts and the spoken line came out as written; one action and one "no other text" line did not.

The Wan 3.0 prompt structure

Alibaba's Wan 3.0 prompt guide (updated October 16, 2026) orders the parts like this: overall description, reference citation, a block for each shot, dialogue, sound and BGM, style, then the negative list. A single sentence already produces a video; the structure is what makes the result match what you pictured.

• For one take, put "Generate single shot" on the first line. For several shots, number them with start and end times, 2–5 seconds each, end to end with no gaps or overlaps. [0-3s] and (00:00-00:03) both work, so pick one and keep it, and name the transition, such as "hard cut".
• Write dialogue as speaker, a speaking verb, a colon and the line in double quotes, then add "Lip sync." for mouth movement. Write "No dialogue." when you want none; otherwise the model decides. Alibaba's older general Wan guide says lip sync to exact words is unreliable, but its examples are from Wan 2.2 and 2.5 (what changed since is in Wan 3.0 vs Wan 2.7). The 3.0 guide adds the "Lip sync" tag, and clip 3 got its line word for word.
• Music has three levels: say nothing and let the model choose, name the music, or write "No background music." Sound effects follow the action, material and weather you describe.
• There is no negative_prompt parameter for wan3.0-video, so negatives are prompt text: a short list of only what must not appear, without padding or repeating the positive prompt. Examples from Alibaba: "No face distortion", "No subtitles", "No costume changes".
• A prompt can be Chinese or English, up to 20,000 characters; longer text is cut off without an error. Alibaba publishes no recommended length, and our five prompts ran 398 to 670 characters.

Alibaba also ships a prompt-tuning Skill for the console, /wan3-pe, which we left alone.

How we made the clips

• Model wan3.0-video through OmniaKey's asynchronous media API on October 27, 2026 (Beijing time), at resolution 720P, leaving prompt_extend, audio and seed at Alibaba's defaults: prompt rewriting on, audio on, a random seed. Clips 1, 2, 3 and 5 were text_to_video, clip 4 image_to_video; ratio was 16:9 except clip 5 at 9:16, duration 10 for clip 2 and 5 for the rest.
• Clip 4 starts from a still we made with GPT Image 2 at 1536×1024, quality medium, cropped to 16:9 and sent in input.media with type first_frame, as an inline JPEG 1024 px wide to stay under the 1 MiB limit.
• Every task succeeded first time: MP4, H.264, 30 fps with an AAC track, 1280×720 (clip 5: 720×1280), 5.04 s (clip 2: 10.03 s). Generation took 197.8 s, 334.3 s, 215.5 s, 153.2 s and 193.8 s per clip, so about 2.5 to 5.5 minutes.
• We viewed frames at fixed times at full size, ran a scene-cut detector on every clip, and checked each soundtrack with two Gemini audio models and a loudness trace; the sound claims below are ones those checks agreed on.
• OmniaKey sells API access to Wan 3.0, so we ran the clips through it. The prices below are Alibaba Cloud's list prices, not ours.
• Sources: Alibaba's Wan 3.0 prompt guide, the Wan 3.0 video generation guide, the API reference and the Model Studio pricing page, checked October 27, 2026.

We ran one clip per prompt, so treat each result as one sample.

Five prompts with real clips

1. One shot: camera, lens, sound and style lines

Clip 1 shows the parts inside one continuous shot.

Generate single shot. One continuous shot, no cuts.
A hand-thrown ceramic teapot sits on a worn oak table beside a rain-streaked window at dawn. A woman's hand lifts it and pours tea into a small white cup; steam curls up through a beam of cold blue window light.
Camera: slow push-in from a medium shot to a close-up of the cup, 85mm telephoto, shallow depth of field, soft background bokeh.
Sound: rain tapping on the glass, tea splashing into the cup, a faint clink of porcelain.
Style: quiet and photoreal, muted teal and warm amber tones, calm morning mood.
No dialogue. No background music.
No text, no subtitles, no watermark.

Settings: text_to_video, 16:9, 720P, 5 s, defaults on.

Result: one unbroken shot in the right mood; the push-in stops short of the cup alone.

• One continuous shot with no cuts: the pot, the white cup and the worn table are there, with out-of-focus rain drops on the window.
• A hand comes in from the right, lifts the pot and pours amber tea while steam rises.
• The push-in lands, though it ends on cup and spout together rather than the cup alone, and the frames show no distinct beam of light. Sound: only the pouring; neither audio model heard the rain.

What to change:

• For a tighter ending, try naming the last framing: "ends on an extreme close-up of the cup".
• Try naming fewer background sounds, tied to something visible.

2. Three timed shots with hard cuts

Clip 2 is the multi-shot test: three numbered shots in 10 seconds, with named transitions.

A 10-second, three-shot sequence of a cyclist at sunrise on a coastal road. Keep the same rider, red jacket and silver bike in every shot.
Shot 1 [0-3s]: wide shot, fixed camera. The cyclist rides along a cliff road above a calm sea, small in the frame, low golden sun.
Shot 2 [3-6s]: hard cut. Tracking shot beside the rider at wheel height, the chain and spinning spokes in sharp focus, the road blurred.
Shot 3 [6-10s]: hard cut. Close-up from the front: the rider exhales and smiles, sunlight flaring behind the helmet.
Sound: wind, tyres humming on asphalt, distant gulls.
BGM: light, hopeful acoustic guitar.
No dialogue.
No text, no subtitles, no costume changes.

Settings: text_to_video, 16:9, 720P, 10 s, defaults on.

Result: all three shots and both cuts landed as written.

• The cut detector found them at 2.97 s and 5.97 s, against the 3 s and 6 s marks.
• Shot 1 is a wide, fixed view of a coastal road at sunrise with the sun low over the sea and a gull in the sky, and the cyclist in red rides from the middle distance until leaving on the left.
• Shot 2 tracks low beside a silver road bike, chain, gears and pedalling legs sharp, road and wheels blurred; shot 3 is a front close-up of a woman in a red jacket and white helmet, mouth open as she breathes out, then a smile as the sun flares behind her.
• Sound: a soft plucked-string melody, wind and gulls, no speech, and the audio models disagreed on the instrument.

What to change:

• If the instrument matters, try describing the playing too: "solo acoustic guitar, slow strumming".

3. A spoken line in quotes, with lip sync

Clip 3 puts one line in double quotes and adds the "Lip sync" tag.

Generate single shot. Medium close-up, fixed camera, eye level.
A barista in her thirties with short curly hair and a green apron stands behind a café counter, holding a takeaway cup. She looks into the camera, smiles and says: "Your oat latte is ready. Have a good day!" Lip sync.
Sound: soft café chatter in the background, the hiss of a milk steamer.
No background music.
Style: natural daylight, warm and friendly, photoreal.
No subtitles, no on-screen text, no face distortion.

Settings: text_to_video, 16:9, 720P, 5 s, defaults on.

Result: the line, the lip movement and "no subtitles" all held.

• One fixed-camera shot: a woman with curly hair, a white shirt and a dark green apron stands behind a counter, holding a takeaway cup with both hands.
• She looks into the camera, smiles and speaks with her lips moving, and both audio models transcribed the line word for word, in two phrases at about 1.0–2.5 s and 3.3–4.0 s.
• The frame stayed clear of subtitles and on-screen text. Sound: café chatter only, and neither model heard the milk steamer.

What to change:

• For two speakers, try giving each line its own sentence and naming the speaker, which is Alibaba's format.
• If you need the steamer, try showing it: "the steam wand hisses beside her".

4. Image-to-video: let the still set the look

Clip 4 starts from a still, so the prompt carries only motion, camera and sound. We made the still with GPT Image 2 and cropped it to 16:9; its prompt asked for a photoreal sunrise shot of a vintage red-and-cream hot-air balloon tethered on a dewy meadow, two passengers in the basket, low mist and a distant tree line. Our GPT Image 2 prompt guide covers the still.

Image 1 is the first frame.
The tether is released and the balloon lifts slowly off the meadow. The burner flares orange twice. The two passengers wave down at the field.
Camera: slow tilt up, following the balloon as it rises.
Sound: the roar of the burner, birdsong, wind.
No dialogue. No background music.
Keep the balloon's colours and the passengers' clothes unchanged. No new people, no text.

Settings: image_to_video, 16:9, 720P, 5 s, still sent as the first frame.

Result: the look and the camera move match; one action was skipped.

• The clip opens on the still's framing with the balloon's colours, the passengers and their clothes unchanged, and no one new. The balloon lifts off and climbs, and the camera tilts up until only sky is left; both passengers raise their arms and wave.
• The flame lights at about 0.7 s and stays lit to the end, so the frames show a single flare; two ropes stay tied to the ground, so the release was skipped.
• Sound: a low roar with no speech or music, loudest in two bursts at about 0.5–1.5 s and 3.5–4.8 s, so the "twice" showed up in the sound; neither model heard birdsong.

What to change:

• Show the release as a visible action: try "a crew member unties the ropes and the ropes fall away".
• For a flame that flares twice, try "the flame flares, dies down, then flares again".

5. Vertical video and sign text in quotes

Clip 5 runs the vertical format and puts the on-screen words in quotes.

Vertical 9:16. Generate single shot.
A narrow back street at night after rain. Above a small noodle shop, a pink neon sign reads "OPEN LATE" in bold capital letters, reflected in the wet pavement. Steam drifts out of the doorway.
Camera: slow tilt down from the neon sign to the shop's doorway.
Sound: light rain, the hum of the neon, distant traffic.
No dialogue. No background music.
Style: cinematic night photography, 35mm film look, magenta and teal.
No other text, no subtitles, no watermark, no logos.

Settings: text_to_video, 9:16, 720P, 5 s, defaults on.

Result: the sign and the tilt match; extra text appeared.

• A vertical clip with a pink neon sign reading "OPEN LATE" on a brick wall above a shop door, spelled right in every frame we checked.
• The tilt works: the sign starts high and is cut off at the top by the end, with the doorway and pink-lit wet pavement filling the frame.
• Two slips, both text: a second, smaller "OPEN LATE" sign in the shop window and marks on the door frame, despite "No other text". Sound: no speech, no music; the audio models disagreed, one hearing water, the other wind.

What to change:

• Describe what you do want as well as what you don't: try "plain dark glass in the window", because the negative list is only for what must not appear.

Common failures and fixes

Alibaba groups its negative lists by the failure they prevent: faces, text, style, sound, action, camera, space and consistency.

What a clip costs

Alibaba charges wan3.0-video by the second of output, at a rate set by the resolution you ask for.

These are Alibaba Cloud's international list prices, checked October 27, 2026; the pricing page marks all three as limited-time 30% off.

wan3.0-video-prime, Alibaba's high-speed version with the same capabilities, lists at $0.068, $0.14 and $0.28 per second for 480P, 720P and 1080P.

The default resolution is 1080P, so a request without it is billed at 1080P: $1.00 for 5 seconds, $0.50 at 720P. Audio costs the same on or off, and it is on by default.

Our five clips came to 30 seconds of 720P, $3.00 at list and $2.10 with the discount. For readers weighing models, see Kling 3.0 vs Wan 3.0; for OmniaKey's current rates, see the Wan 3.0 model page.

Run these prompts through the API

This request sends clip 2's cyclist prompt, with the shot lines as \n inside the JSON string and a fixed Idempotency-Key:

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: cyclist-shots-001" \
  --data-binary @- <<'JSON'
{
  "model": "wan3.0-video",
  "type": "video",
  "operation": "text_to_video",
  "input": {"prompt": "A 10-second, three-shot sequence of a cyclist at sunrise on a coastal road. Keep the same rider, red jacket and silver bike in every shot.\nShot 1 [0-3s]: wide shot, fixed camera. The cyclist rides along a cliff road above a calm sea, small in the frame, low golden sun.\nShot 2 [3-6s]: hard cut. Tracking shot beside the rider at wheel height, the chain and spinning spokes in sharp focus, the road blurred.\nShot 3 [6-10s]: hard cut. Close-up from the front: the rider exhales and smiles, sunlight flaring behind the helmet.\nSound: wind, tyres humming on asphalt, distant gulls.\nBGM: light, hopeful acoustic guitar.\nNo dialogue.\nNo text, no subtitles, no costume changes."},
  "parameters": {"duration": 10, "resolution": "720P", "ratio": "16:9"}
}
JSON

1. The reply is 202 with a task id.
2. Fetch https://api.omniakey.com/v1/media/tasks/{id} with the same id and loop on status, not progress.percent, which can stay at 0 while the clip renders. Our clips took 2.5 to 5.5 minutes: keep polling and do not resubmit, since the same Idempotency-Key and body return the same task and are not charged twice.
3. Download the clip from assets[0].url, a signed link that lasts about an hour, and poll again for a fresh one if it expires.

For image-to-video, set operation to image_to_video and send the still in input.media with type first_frame, as an HTTPS URL or a data URI, whole request at most 1 MiB, and start the prompt with "Image 1 is the first frame." For the faster tier send wan3.0-video-prime. A cancelled Wan 3.0 job on OmniaKey keeps rendering and is billed if it succeeds, because Wan 3.0 has no provider-side cancel. Start from the Wan 3.0 model page, and see the video API docs.

Common questions

How long can a Wan 3.0 prompt be?

The API takes up to 20,000 characters, counting each Chinese character or letter as one, and cuts anything longer without an error. Alibaba publishes no recommended length; our five prompts ran 398 to 670 characters.

Does Wan 3.0 support negative prompts?

There is no negative_prompt field for wan3.0-video; negatives are a short list at the end of the prompt text. Clip 5 added a second, smaller "OPEN LATE" sign in the shop window despite "No other text", so describe the glass you want as well.

Should I turn off prompt_extend?

prompt_extend defaults to true, and an LLM rewrites the prompt before generation, which helps short prompts and adds latency. Alibaba gives no guidance on switching it off for long structured prompts, so we left it on, and clip 2 still cut at 2.97 s and 5.97 s.

Can Wan 3.0 make a 30-second video from one prompt?

duration takes 2 to 30 seconds, so one prompt can fill 30 seconds. Alibaba's long examples are built from 2–5 second shots, which is how clip 2 reached 10 seconds.

Can I write Wan 3.0 prompts in Chinese?

Yes. The API accepts Chinese and English prompts, and the cap counts each Chinese character as one. We wrote ours in English.

Update log

• October 28, 2026: first version, with five prompts run on October 27, 2026.