Wan 3.0 Prompt Guide: Reference Syntax and Sound Direction

There is no reference syntax to learn for Wan 3.0 on VdoBloom. The model’s multimodal reference inputs exist upstream but are not exposed on the platform yet, so a Wan 3.0 prompt does not address numbered images or tagged references — it describes one continuous shot in plain prose. Sound direction works the same way: audio is generated natively in the same pass, there is no dialogue field, voice picker or music selector, and the prompt is the only lever you have over what the clip sounds like.

Updated October 2026.

Disclosure: VdoBloom is our platform.

What you actually control

A Wan 3.0 generation on VdoBloom’s text-to-video tab has five inputs, and only one of them is free text:

Duration — any whole second from 2 to 30.
Resolution — 480p, 720p or 1080p.
Aspect ratio — 16:9, 4:3, 1:1, 3:4 or 9:16.
Model variant — Wan 3.0, or Wan 3.0 Video Prime for faster generation.
The prompt — between 2 and 20,000 characters, covering everything else: subject, action, camera, light, setting and sound.

That last item carries a lot of weight. Because there is no separate audio field and no reference-slot UI, the prompt is where you describe not just the picture but the sound that should come with it. A prompt that only describes what the camera sees tends to get generic ambience; a prompt that names the room and the objects making noise gets something you can use.

Reference syntax: what is not available

If you have used reference tags or numbered image slots on other models, the instinct is to look for the same thing here. Wan 3.0 supports multimodal reference inputs upstream — video, audio, link and file references — but they are deliberately not exposed on VdoBloom yet, because they change how a generation is billed: pricing moves from output seconds to input seconds plus output seconds, which needs its own price tier before it can be offered honestly.

So for now, a Wan 3.0 prompt should never contain a phrase like “use image 2 for the face”. Nothing will read it. What you can do instead:

Describe the thing in words. If a reference existed, what would you be pointing the model at? Describe that object’s shape, colour, material and placement instead.
Start from an image on the image-to-video side. Wan 3.0 is an all-in-one model that takes text or a source image, so supplying a start image is the supported way to anchor the first frame — not a reference tag inside the prompt.
Use a different model for pinned frames. If you need an exact final frame, Wan 2.7 Image-to-Video keeps a first- and last-frame control that Wan 3.0 does not list, and that is the cleaner tool for a planned transition.
Look elsewhere for tagged references. Numbered reference images are documented on the platform as a Seedance 2.5 feature, not a Wan one.

Sound direction: writing audio as prose

Wan 3.0 generates audio in the same pass as the picture, at no separate charge, and there is no audio switch to find. The practical consequence is that sound is a writing problem. Four habits make it work:

Name the source, not the mood. “Tense ambience” produces nothing specific. “A fridge humming in the next room, a clock ticking, footsteps on lino” gives the model objects to make noise with.
Place the sound in the room. Distance and surface change what a sound is: close birdsong and distant traffic, or rain on glass rather than rain in the open.
Say what continues and what changes. For a long take, tell it whether the sound should stay constant or shift with the action — “the engine note rises as the car pulls away”.
Keep voices light in the frame. There is no dialogue field or voice picker, so treat speech as part of the scene description rather than a script. If a clip needs exact spoken lines, plan to record or generate the voice track separately and lay it over the finished clip.

A structure that holds up

The prompts that render reliably on a 15- or 30-second timeline share an order: subject, action, camera, light, setting, sound. Long prompts do not help unless each sentence adds information the model can act on.

Subject. Who or what, with the one or two details that must be right.
Action. What changes across the clip. A single continuous action is easier to hold than a sequence of events.
Camera. Static, push in, pull out, orbit, tilt, handheld. One camera behaviour per generation.
Light. Direction and quality — golden hour from the left, hard overhead sun, soft window light.
Setting. Enough detail to fix the place, without a second scene’s worth of props.
Sound. Two to four named sources, plus whether the bed stays constant.

Six prompts to work from

Dialogue-free character study, 10s. “A single continuous take: an older man in a grey coat sits on a bench, opens a newspaper and reads. Static camera, 50mm, soft overcast light, a quiet park behind him. Sound: birdsong close, distant traffic, paper rustling as he turns a page, no music.”
Motion that develops, 30s. “A single unbroken take of a runner leaving a city side street and reaching a bridge. Camera tracks alongside at jogging pace, gradually rising. Wet asphalt, low grey light, sodium street lamps still on. Sound: footsteps on wet pavement, breath, a bus passing behind, wind picking up towards the end.”
Product, 6s. “A glass bottle of amber liquid on a dark slate surface. Slow orbit from the left, hard key light raking across the label, tight framing, shallow depth of field. Sound: low room tone, the soft clink of the bottle being set down.”
Image start, 12s. “Begin on the supplied photograph and hold the composition for the first two seconds, then tilt slowly upward to reveal an approaching storm. Keep the building’s architecture exactly as in the still. Sound: wind building, a shutter creaking, thunder far off.”
Interior with shifting ambience, 20s. “A bakery before opening. A baker moves between the counter and the oven, pulling trays. Camera drifts forward at knee height. Warm light from the oven, cool light from the front window. Sound: the oven door opening and closing, trays on metal, a radio in the back, the ambience brightening as a door opens at the end.”
Cheap test shot, 2s. “A hand turns a brass tap, tight framing, hard side light. Sound: a single splash of water.” At 480p this is 7 credits, which is the point — find out whether the motion reads before spending on a long take.

Common mistakes

Writing a scene list. Timestamped, shot-by-shot prompts are documented on VdoBloom for the Seedance 2 family, not for Wan. A Wan 3.0 prompt describes one continuous shot; describing four cuts produces a muddled take.
Leaving sound to chance. A silent description gets generic ambience. Two to four named sources is the difference between a clip with a room in it and a clip with a hum.
Reusing reference-tag habits. Phrases that address image two or a tagged reference do nothing here.
Buying a long render first. Test at 2 seconds and 480p, where a run is 7 credits, then climb to the duration and resolution you actually need.
Paying for Prime while writing. Prime is roughly 55 to 60 percent more per second than the base model for identical specs; it belongs on a final pass, not on a draft.

Which model to pick

Need	Model	Cost signal
Long single take with sound, lowest cost	Wan 3.0	About 3.2 credits a second at 480p, 6.4 at 720p, 12.8 at 1080p
Same shot, faster turnaround	Wan 3.0 Video Prime	About 55 to 60 percent more per second
An exact first and last frame	Wan 2.7 Image-to-Video	Caps at 15 seconds; keeps frame control Wan 3.0 does not list
Timestamped multi-shot prompts and numbered references	Seedance 2.5	A different family; priced per second and per resolution

Costs are shown on the generate button before every run, and the one-second duration ladder means you can choose a 7-second clip rather than rounding to 10 — worth doing, since sound and picture scale linearly with length.

Frequently asked questions

Does Wan 3.0 support reference images in the prompt?

No. The multimodal reference inputs exist upstream but are not exposed on VdoBloom, because they change billing to input seconds plus output seconds and need their own pricing tier. Describe the object in words, or anchor the starting frame by supplying a source image on the image-to-video side.

How do I control the audio?

By describing it. Audio is generated natively in the same pass, there is no dialogue field, voice picker or music selector, and no audio toggle. Name the sources, place them in the room and say whether the bed stays constant or changes with the action.

Can I write a multi-shot prompt with timings?

Not productively. Timed, shot-by-shot prompt formats are documented on the platform for the Seedance 2 family. Wan 3.0 renders one continuous shot, and a prompt that lists several scenes tends to compress them into a muddle.

Why is my clip’s sound generic?

Because nothing in the prompt asked for anything specific. Add two to four named sources and a note on distance — close footsteps, traffic behind, rain on glass rather than rain in the open — and the sound stops being wallpaper.

How many credits should a prompt test cost?

Seven. That is a 2-second 480p generation, the lowest-cost run available, and it is enough to see whether the subject and motion behave before you pay 32 credits for 5 seconds at 720p or 384 for 30 seconds at 1080p.

Start with a two-second test on Text to Video, add sound sources only once the picture holds, and extend the duration a few seconds at a time. Common prompt mistakes covers the wording traps in more detail, and the pricing page has the plan and credit-pack details if you are budgeting a longer project.