Wan 3.0 prompt guide

The formula is from Alibaba’s Wan 3.0 prompt guide. For the model, see the Wan 3.0 page.

Alibaba says one sentence is enough for Wan 3.0, and the more precise the description, the closer the result. Its full formula has seven parts, in one paragraph; leave out any you don’t need.

1. 1Open with one line on the theme, the point of view, the length and the pacing.
2. 2If you upload pictures, say what each is for: “the girl in Image 1 stands in the room of Image 2”.
3. 3Write each shot with its seconds: the subject, the scene, the motion and the camera.
4. 4Add the dialogue as Name says: “…”, then the sound effects and the music, or “No background music”.
5. 5Finish with the art style, color tone and mood, then a short negative list.

Wan 3.0 can tell a story in several shots inside one clip. Alibaba’s format is a number, the seconds and the content: “Shot 1 [0-3s] …”. The shots meet end to end, with no gaps or overlaps, at 2 to 5 seconds each; its API guide suggests 4 to 6.

On DreamBench, a Wan 3.0 clip doesn’t cut to a new setting partway through: each clip is written as one continuous shot, and a change of place comes at the next clip.

Alibaba asks for camera moves in plain words and allows a lens and depth of field. Put the move inside the shot it belongs to.

Alibaba warns against actions faster than the model can control, such as “shaking head 3 times per second”. Give a motion a pace a person could act.

Wan 3.0 makes speech, effects and music in the same pass as the picture. See native audio.

• Dialogue: the character, a speaking verb, a colon and the line in quotes: she whispers: “Not yet.”
• Several speakers: one line each, in order, each with its speaker.
• Narration: Voiceover: “Every morning starts here.”
• Silence: write “No dialogue”, or the model decides by itself whether anyone speaks.
• Effects: describe what makes the sound, such as rain on a tin roof or boots in a puddle, and Wan adds it; or give an effect its own sentence and time: “From the 8th second, a church bell.”

Alibaba’s Wan 3.0 pages list no dialogue languages; its examples speak English and Chinese. Name the language before a line, and check each take. See lip sync.

Alibaba groups the look into three parts, art style, color tone and mood, said at the start or the end: “clay animation, high-saturation warm yellow, light and healing”.

• List only what you don’t want, and leave the list empty if there is nothing.
• Don’t pad it or repeat what the prompt already says.
• Typical lines: no subtitles, no watermark, no extra fingers, no face distortion, no frequent shot cuts.
• For an animated style, rule out the other look: “no photorealistic live-action”.

Alibaba’s platform takes 2 to 30 seconds, or a smart length the model picks. Write as much as the seconds hold:

In a DreamBench film each Wan 3.0 clip runs 3 to 12 seconds, as long as its action takes, and DreamBench writes its prompt to fit.

Wan 3.0 can start from a first frame, or run between a first and a last frame. Alibaba says the picture already fixes the subject, the scene and the style, so the prompt describes the motion, the plot, the dialogue and the camera. See first and last frame.

1. 1Give the picture as the first frame, or the first and the last.
2. 2Describe what happens next, shot by shot with times if there are several, without repeating what the picture shows.
3. 3Add the lines and the sound; for a last frame, describe the change that leads to it.

On Alibaba’s platform, first and last frames can’t be mixed with reference pictures in one request. On DreamBench, every Wan 3.0 clip starts from an opening frame made from your references. See image to video.

On Alibaba’s platform, uploads are numbered in order, with images, videos and audio counted separately: Image 1, Video 1 and Audio 1 can all exist. Alibaba’s advice is to give each a purpose, so features don’t bleed between people.

• Lock each subject to one upload: “the little girl is Image 2, the shop is Image 1”.
• Describe each character in the same words in every shot and every clip.
• Keep the clothes, the props and the light from shot to shot.

On DreamBench, each character, place and prop is drawn once as a sheet, and every clip’s opening frame is made from the sheets it needs. See character consistency and reference to video.

On Alibaba’s platform one Wan 3.0 clip can hold eight timed shots. On DreamBench, a film is several clips, each one continuous shot from its own opening frame, joined at hard cuts, so any one can be redone.

Describe the film on DreamBench and it writes every clip’s prompt, draws the cast and has Wan 3.0 animate each clip. See multi-shot video.

Each follows Alibaba’s order. Paste one into Alibaba’s platform, or use it as a DreamBench brief; change the subject, keep the shape.

One continuous shot: the harbor at dawn

One continuous shot. A fishing harbor at dawn, calm and quiet, 10 seconds. The camera glides low over the water toward a red boat as a fisherman in an orange oilskin coils a rope and looks up at the sky. Gulls, water lapping at the hull, a distant engine. No dialogue. No background music. Documentary realism, cool blue tones warming to gold, peaceful.

A three-shot drama

A short drama about forgiveness, 12 seconds, pacing from tense to warm. Shot 1 (00:00-00:04): medium shot, fixed camera; a young man stands at a farmhouse door, holding a cake box. Shot 2 (00:04-00:08): hard cut, close-up, 85mm telephoto; his father opens the door, unsmiling. Shot 3 (00:08-00:12): the father steps aside and says: “Your mother kept your seat.” Wind, a dog barking far off. 35mm film look, muted greens, tender. No subtitles.

Two speakers in a lift

One continuous shot, 8 seconds, light office comedy. Inside a mirrored lift, a woman in a navy suit and a courier with a tall stack of boxes. Fixed shot, camera static. The courier says: “Which floor?” She glances at the boxes and replies: “All of them, apparently.” The lift hums and pings. No background music. Bright, clean light, natural color.

A product with a voiceover

A product film for a steel water bottle, 10 seconds, calm and confident. Shot 1 (00:00-00:05): the bottle on wet rock by a waterfall, mist drifting, the camera orbiting slowly. Shot 2 (00:05-00:10): a hand fills it from the stream and closes the cap with a click. Voiceover: “Cold for a day. Built for a lifetime.” Rushing water, the cap’s click. Soft acoustic guitar under the voice. Clean commercial look, cool greens.

Sport with no dialogue

One continuous shot, 8 seconds. A sprinter in a white tank top bursts out of the blocks on a wet track under floodlights and runs toward the camera. Handheld follow at running pace, low angle. Spikes on the track, her breathing, rain. No dialogue. No background music. Realistic sports-broadcast look. No slow motion.

A music cue: rooftop violinist

One continuous shot, 15 seconds, dreamy and romantic. A violinist in a long gray coat plays on a Paris rooftop at dusk, chimneys and the city behind her. The camera pulls back and rises slowly. A slow solo violin melody that builds as the lights of the city come on. Wind in her hair. No dialogue. Warm pink and blue dusk tones.

Clay animation

Clay animation, high-saturation warm yellow, light and healing, 10 seconds. Shot 1 (00:00-00:05): a small clay bear in a striped scarf waters a single sunflower on a windowsill. Shot 2 (00:05-00:10): the flower blooms in three soft bounces and the bear claps. Squelchy clay footsteps, a happy chime. No dialogue. No photorealistic live-action, no extra limbs.

From a first frame

Continue the market scene from the first frame, 8 seconds, one continuous shot. The spice seller lifts a scoop of saffron, lets it fall back into the sack and smiles at a customer, who says: “Two hundred grams, please.” The camera pushes in slowly. Market chatter, a scale clicking. No subtitles.

First and last frame: a sketch that comes to life

One continuous shot, 10 seconds, from a pencil sketch of a horse (first frame) to a real horse galloping in a field (last frame). The pencil lines thicken, color floods in from the edges, and the horse breaks into a gallop as the paper becomes grass. The camera pulls back slowly. Pencil scratching, then hooves and wind. No background music.

Two references

A warm bakery morning, 10 seconds. The baker’s look follows Image 1; the bakery follows Image 2. Shot 1 (00:00-00:05): fixed shot, she pulls a tray of croissants from the oven. Shot 2 (00:05-00:10): close-up, she breaks one open and steam rises; she says: “Perfect.” Crackling pastry, the oven door. No background music. Golden morning light. No extra people, no costume changes.

A 30-second story

A comic short about a lost umbrella, 30 seconds, brisk build-up and a warm ending. Shot 1 (00:00-00:05): wide shot, rain on a city square; a man’s umbrella blows away. Shot 2 (00:05-00:10): handheld follow as he chases it past a café. Shot 3 (00:10-00:15): it lands at the feet of a girl in a yellow raincoat. Shot 4 (00:15-00:20): she holds it out and says: “Is this yours?” Shot 5 (00:20-00:25): he nods, soaked. Shot 6 (00:25-00:30): they share it; the camera pulls back and rises. Rain, wind, a light jazz piano. Bright, saturated color. No subtitles, no watermark.

Most weak Wan prompts leave a part to chance: the seconds, the speaker or the sound.

• One prompt run twice gives two clips.
• Alibaba says Wan 3.0’s audio texture and the accuracy of on-screen text are still improving.
• One generation stops at 30 seconds.
• A real person or product can’t be known from words; it takes a picture you have the right to use.
• On DreamBench, Wan 3.0 always makes sound, and its clips don’t cut to a new setting mid-clip.

• A film: describe it, add up to 8 reference pictures and pick Wan 3.0. DreamBench writes each clip’s prompt, makes its opening frame, has Wan 3.0 animate it with sound and joins the clips; redo any one with a note.
• Settings: 1:1, 16:9, 9:16 or 4:3, 720p or 1080p, on every plan.
• Not offered: the Video page, which takes models that start from words alone; Alibaba’s numbered references, documents, editing and extension; a silent clip.

The AI video prompt generator writes a prompt in this order for any model. The price in credits shows before you generate; the first time, create a free account and choose a plan. For a prompt sent as typed, see the Seedance 2.5 or Kling 4.0 guides; for prompts from words alone, text to video.