ChatGPT
Perplexity
Claude
Gemini
Wan 3.0 is one of the newest AI models on OpenArt's AI video generator, and Alibaba published its own prompt formulas the day the model launched. Almost nobody reads them, so we've put together this definitive prompt guide for you.
Wan 3.0 makes up to 30 seconds of video at once, takes 20 reference files, and generates audio at the same time as the visuals. This guide covers the formulas the makers of Wan, Alibaba, wrote down, plus the three things Wan 3.0 decides for you automatically if you don't specify them.
Key takeaways
- Start with a subject, a place, and a movement. Add the camera, the style, and the sound once that works.
- Label every reference file with @Image1, @Video1, or @Audio1 so Wan 3.0 knows what each one is for.
- Number your shots and give each one a time range like [0-6s].
- Write voice lines, sound effects, and music as three separate things.
- Say "no background music" when you do not want it, because Wan 3.0 adds music and voice lines on its own.
- Switch Auto Polish off once your prompt is exact, so it does not reword your labels and time ranges.
Start with three things, then add more
Alibaba's starter formula is short. You need a subject, a place, and a movement. "A red fox walks across a snowy field at sunrise." That is all three, and it is enough to get a real video back.
Once that works, the fuller formula adds three more: how the shot looks, the style, and the sound. Add them one at a time so you can see what each one changed. You have plenty of room, because a Wan 3.0 prompt can run to 20,000 characters.
Label every reference with @ tags
Wan 3.0 takes up to 20 reference files, and OpenArt counts them for you at the top of the upload box.
Uploading them is only half the job. Wan 3.0 will not guess that picture 3 is the coffee shop and picture 4 is the jacket. You have to say so in the prompt, using the @ label:
@Image1 is the woman. @Image2 is the coffee shop. @Audio1 is her voice. The woman from @Image1 sits by the window in @Image2 and orders a coffee.
Alibaba's own examples label every single reference this way. You can also name the same reference more than once in one prompt, which is how you control where it applies.
Start and end frames work the same way. On OpenArt you can set the picture your video opens on and the picture it closes on, and still add reference files next to them. Say in the prompt which picture is the opening frame, which is the closing frame, and what each reference is for.
Your video references can total 15 seconds, and your audio references another 15 seconds. If you want the same person in video after video, save them once in OpenArt Characters and reuse them instead of hunting for the original pictures every time.
Number your shots and give each one a time
Wan 3.0 makes up to 30 seconds at once, so tell it what happens across that time instead of describing one frozen picture. Alibaba's format is a short summary first, then numbered shots with time ranges:
A short story about a girl who finds a stray cat.
Shot 1 [0-5s]: She walks home in the rain and stops under a doorway.
Shot 2 [5-12s]: Close up on a small wet cat looking up at her.
Shot 3 [12-20s]: She kneels down and holds out her hand.
Give each range one clear thing to do. If a shot feels rushed in your first result, widen that one time range and leave the rest alone. For a longer story with more shots than this, build it out scene by scene in OpenArt Director.
Write the sound as three separate things
Wan 3.0 makes the sound at the same time as the picture, so you can shape it with words. Alibaba splits sound into three parts, and it helps to write each one separately.
Voice lines. Say the line, then say how it is said. "She says, 'I told you this would happen.' Fast, a little annoyed, American English."
Sound effects. Name the thing, what it does, and the background noise. "A glass falls off the table and shatters on the wood floor, quiet room."
Music. Name the music and the style. "Slow piano under the whole scene."
Three things Wan 3.0 decides for you
This is the part that catches people out. When you say nothing, Wan 3.0 still makes a choice.
Voice lines. Write out what someone says and Wan 3.0 keeps your exact words. Say nothing and it will make up voice lines of its own. Turn that off with "no voice lines in the whole video."
Music. Wan 3.0 picks background music based on what your video is about. Turn that off with "no background music."
Number of shots. Wan 3.0 decides on its own whether to cut between shots. Ask for one unbroken shot with "one continuous take."
Length and shape work the same way. Leave them blank and Wan 3.0 picks a length between 2 and 30 seconds, plus an aspect ratio that suits your scene.
Check four controls before you generate
On OpenArt, four settings sit around the prompt box and each one changes what your words do.
Auto Polish. This rewrites your prompt before it runs. That helps when you have typed one loose sentence, but it can also reword the reference labels and time ranges you set on purpose. Switch it off once your prompt is exact.
Audio. Leave this on and Wan 3.0 makes sound with the picture. Switch it off for a silent clip, which is the quickest way to stop unwanted music.
Camera. Pick your camera move here instead of typing it, and you get the same move every time without wording it differently in each prompt.
Mode. Standard and Prime set the quality level. Test your prompt on Standard, then move the version you like to Prime.
You can also queue up to 8 videos at once. That is the fastest way to learn what a word is doing: change one thing per version, run all 8, and compare.
Say less when you start from a picture
When you upload a starting picture, that picture already settles who is in it, where they are, and how it all looks. Repeating that in the prompt just fights the picture you uploaded.
So write two things only: what moves, and what the camera does. "She turns toward the window and smiles. Slow push in." If you want the camera to hold still, say "fixed shot," because Wan 3.0 will move it on its own otherwise.
Fix one thing instead of making the video again
Wan 3.0 edits video from plain instructions, so one small problem does not cost you a whole new video. Name what you want changed, then say the rest stays the same.
Remove the sunglasses. Her movement and everything else in the video stays the same.
The same wording works for adding something, changing the style, fixing the light, or swapping a voice line.
Four mistakes that waste generations
- Describing a picture instead of a video. Say what changes across the 30 seconds.
- Leaving references unlabeled. If you upload six files, say what each one is for.
- Stacking camera moves. One clear move beats three fighting each other.
- Forgetting to turn off sound you never asked for. Music and voice lines show up on their own.
Frequently asked questions
How many references can Wan 3.0 use?
Wan 3.0 takes up to 20 references in one video: 10 images, 5 video clips, and 5 audio clips. You can add one document or web page on top of that, so a character, a product, a location, and a voice can all guide the same shot. Your video references can total 15 seconds, and your audio references another 15 seconds.
How long can a Wan 3.0 video be?
Any length from 2 to 30 seconds. If you leave the length blank, Wan 3.0 picks a length that suits your prompt.
Does Wan 3.0 make sound?
Yes. Wan 3.0 makes voice lines, sound effects, and music at the same time as the picture. Describe the sound you want in the prompt, or switch it off if you want a silent clip.
Can I use reference files and a start and end frame at the same time?
Yes. On OpenArt you can set the picture your video starts on, the picture it ends on, and still add reference files alongside them. Label each one in the prompt so Wan 3.0 knows which picture opens the video, which one closes it, and which references carry your character, product, or voice.