Wan 3.0 Prompt Guide: Structure, Camera, References

The prompt formula: subject → place → movement → camera → style → sound

Wan 3.0 prompts read more like a short shot brief than a list of keywords. Several independent guides converge on the same order: describe the subject and what it's doing, name the setting, say how the subject moves, then layer in the camera move, the visual style and finally the sound. You don't need every layer for every shot — a one-line subject-and-place prompt is a valid starting point — but for anything with a specific look, working through the order top to bottom keeps the prompt readable and keeps the model from having to guess which detail matters most.

A minimal version might be just "a fox crosses a snowy field at dawn." A fuller version adds all six layers, in the same order the model expects them:

A red fox trots across an open snowy field just after sunrise. It stops, sniffs the air, then continues toward the treeline.
Camera performs one slow push-in from a wide shot to a medium shot, locked-off otherwise.
Soft golden backlight, shallow depth of field, realistic fur detail, cinematic nature photography.
Quiet wind, distant birdsong, no music.

Two details worth keeping in mind: the model auto-fills anything you leave out (it will pick its own duration, camera move, even background music), and one Wan 3.0 prompt can run up to roughly 20,000 characters — far more than you actually need. A tight, well-ordered paragraph beats a long one; once a prompt runs past a few hundred well-chosen words, the model has more to balance and starts favoring some details over others.

On aipixel.fun, the model's limit isn't the only one worth knowing: the generation form (/generuj) and its API don't enforce a prompt-length limit of their own, but a prompt saved to the public library has a hard 5,000-character limit (the database rejects anything longer) — if you plan to publish a prompt, it's worth staying under that from the start.

Duration and output size are worth locking in early rather than leaving to the defaults: Wan 3.0 accepts any whole number of seconds from 2 to 30, and renders at 480p, 720p or 1080p across several aspect ratios (16:9, 9:16, 1:1, 4:3 and 3:4 all show up in the documentation). Deciding on these before you write the rest of the prompt keeps the shot description consistent with the frame you'll actually get.

Camera movement: name the move, keep it to one per shot

Wan 3.0 responds well to camera language borrowed from film sets rather than vague words like "cinematic." The moves that show up consistently across prompting guides are: static / locked-off (no movement), dolly or push-in (camera moves toward the subject), pull-back or dolly out (camera moves away), pan and tilt (camera pivots horizontally or vertically from a fixed point), orbit (camera circles the subject) and handheld or tracking (camera follows alongside, with natural motion). Lens choices like a 35mm or 50mm focal length, or an anamorphic look, are recognized too and shape the framing.

The single most repeated piece of advice across sources: describe one dominant camera move per shot, not three at once. "Slow push-in while orbiting and tilting up" gives the model conflicting instructions to reconcile; "one slow push-in, locked-off otherwise" gives it one clear job. If a 30-second video needs to feel more dynamic, sequence separate moves across separate shots (see the next section) rather than stacking them into a single instruction.

Camera performs one slow, smooth push-in from a medium-full shot to a medium shot. No other camera movement.

References: @Image1, @Video1, @Audio1 and assigning roles

When you attach a reference image, video or audio clip, Wan 3.0 numbers it by the order it was attached and expects you to refer back to it by that number inside the prompt text — for example @Image1 for the first reference image, @Video1 for the first reference video, @Audio1 for the first reference audio clip. A second image becomes @Image2, and so on. Some guides write the label without the @ ("Image 1"); both forms point at the same idea — the label is what lets you tell the model which reference owns which part of the shot.

Labelling matters as much as attaching the file: a bare reference with no explicit instruction leaves the model to guess what it's for. State the role directly — "@Image1 is the main character, preserve the face and hairstyle exactly" — rather than assuming the model will infer it. This is exactly the pattern our own published prompts use whenever a shot needs a consistent face across the video.

@Image1 is the main character. Preserve the face, hair, skin tone and facial features from @Image1 exactly, without alteration..

One continuous unbroken take, no cuts. The same location throughout: a narrow street in Tokyo at night, light rain, glowing neon signs reflecting on the wet pavement.

The person from @Image1 walks slowly toward the camera holding a transparent umbrella...

On the number of references: guidance converges on up to 20 reference assets total across images, video and audio in a single generation, so there is real headroom to combine a face reference, a location reference and a voice reference in one prompt. A reference video clip should stay short — around 15 seconds or less — since the combined length of your references and the requested output has to fit inside the model's own duration ceiling.

Multiple shots and timing: [0-6s] notation vs. one continuous take

For a video that needs to change setup partway through, one documented approach numbers each shot and gives it an explicit time range: "Shot 1 [0-5s]: ...", "Shot 2 [5-10s]: ..." and so on, with one clear action per range. This is the same pattern you'll see in several of our own published prompts below, and it keeps a 10-, 15- or 30-second video from turning into a single confused instruction trying to cover everything at once.

Shot 1 [0-5s]: Close-up, the cat carefully presses dough with its paws, flour dusting into the air.
Shot 2 [5-10s]: Top-down shot, the cat folds a pierogi and pinches the edges with its paws.

The opposite approach — and just as common in practice — is asking for one continuous, unbroken take with no cuts at all: "One continuous unbroken take, no cuts. The same location throughout." This tends to hold a consistent subject, location and lighting more reliably than a multi-shot prompt, at the cost of not being able to jump between distinct setups. Pick shots-with-timing when the idea genuinely needs separate beats; pick one continuous take when it doesn't.

Sound: voice lines, effects, and turning off the music

Wan 3.0 generates audio together with the picture in the same pass rather than as a separate step, so sound is worth writing on purpose instead of leaving to chance. Guidance groups audio prompting into three layers: voice lines (quote the exact line, and describe who's speaking and how), sound effects tied to what's happening on screen, and background music, which the model will add on its own unless you say otherwise.

• Turn off background music — add "no background music" (or "no music") to the prompt.
• Turn off spoken dialogue — add "no voice lines" (or "no dialogue") to the prompt.
• Add a sound effect — name it plainly and tie it to the action — "soft footsteps on wet pavement," "quiet camera shutter clicks" — rather than describing a mood.

Distant traffic and city ambience. No dialogue. No music.

Text-to-video without a reference: when it's enough

Not every shot needs a reference image. When the subject doesn't need to match a specific real person, product or location — a stylised character, an abstract concept, a generic setting — a fully written text prompt following the same subject → place → movement → camera → style → sound order is enough on its own. Several of our own published prompts skip references entirely and rely purely on a detailed text description.

Extreme close-up of a perfectly ripe strawberry resting on a white ceramic plate. A tiny incision appears on its surface and hundreds of miniature translucent red glass cubes slowly pour out from inside. Camera performs an extremely slow lateral macro slide from left to right, 100mm macro lens, razor-thin depth of field.

Reach for a reference image specifically when a recognisable face, product or exact location has to carry over into the output unchanged — that's the one thing a text-only prompt genuinely cannot guarantee.

Common mistakes

• Stacking multiple camera moves in one shot — name one dominant move; sequence the rest across separate timed shots instead.
• Attaching a reference without labelling its role — state explicitly what @Image1/@Video1 is for — face, product, location — don't leave it to be inferred.
• Writing a wall of adjectives instead of an action — describe what happens and how the camera moves; style words support the action, they don't replace it.
• Assuming silence by default — audio is generated automatically unless you say "no background music" / "no voice lines" — silence has to be requested.
• One giant instruction for a 30-second video — split it into timed shots ([0-6s], [6-12s]...) instead of asking one sentence to cover the whole runtime.
• A reference video or audio clip longer than needed — keep reference clips short (around 15s or less) — the combined reference and output length shares the same duration ceiling.

Examples from the library

@Image1 is the main character. Preserve the face, hair, skin tone and facial features from @Image1 exactly, without alteration..

One continuous unbroken take, no cuts. The same location throughout: a narrow street in To…

A fluffy orange tabby cat wearing a tiny white chef's hat stands on a stool at a rustic wooden kitchen counter, cozy Polish grandmother's kitchen, warm afternoon window light, shallow depth of field.

Shot 1 [0-5s]: Clos…

@Image1 is the main character. Preserve the face, hair, skin tone and facial features from @Image1 exactly, without alteration.

Dark apartment kitchen at night, only the fridge light glows. Realistic handheld smartphone…

Generate a single continuous 7-second shot. A pristine transparent glass perfume bottle stands centered on a polished black stone pedestal in a dark luxury studio. At first the bottle is completely still. Tiny golden par…

7-second luxury automotive commercial, one continuous shot. A matte-black futuristic sports car is parked alone on a mirror-like salt flat at sunrise. The camera starts extremely close to the rear wheel and glides smooth…

Sources

• OpenArt — Wan 3.0 prompt guide
• Runware — Wan 3.0 prompting docs
• Morphic — Wan 3.0 model resource
• Venice — Wan 3.0 prompt tips
• Alibaba Cloud — Wan3.0: 30-second AI video generation from any input