Writing Cinematic Prompts That Hold

A prompt Seedance 2.5 actually understands doesn't read like an essay. The same shot can come out as a smooth cinematic long take, or as a character whose face drifts halfway through the clip — and the difference is almost entirely in how the prompt is written. This guide collects a format that holds up across a lot of testing. Copy the structure and your prompts will start looking like the professional ones.

The counterintuitive part: mixing languages on purpose

Camera moves, lighting, texture, and action → write these in English. Terms like slow handheld tracking shot or Kodak 35mm film grain land far more accurately in English than in a translated version — the model's training data simply carries more weight for these terms in English.

Dialogue, voice-over, and any on-screen text → write these fully in your target output language. If your dialogue is in English, write the whole line in English and don't mix languages inside it, and don't expect the model to translate it correctly for you — it usually won't land on the exact line you meant.

Setting, everyday objects, and local, lived-in details → often read more accurately in the local language they belong to. Culturally specific, everyday nouns are frequently understood better when written in the language they naturally belong to rather than translated. Don't force a translation just for the sake of consistency.

So a normal prompt looks like this — mixing terminology and setting language inside one line is the norm, not a mistake:

Whole scene: A quiet afternoon in an old traditional courtyard house, the scent of osmanthus drifting through the yard, shot on 16mm Kodak Portra 400 film, warm soft golden hour lighting, natural film grain, gentle lens bokeh, quiet nostalgic daily atmosphere, no harsh shadows.

Terminology stays in English, setting stays in the language it belongs to — each side does what it's best at.

Four beats, no timestamps

Break a shot into four fixed beats instead of cramming everything into one sentence, and motion continuity holds together noticeably better:

Opening shot: an establishing shot, a quiet start
First action beat: the first move or blocking
Second dialogue beat: dialogue or a second action
Final closing beat: the closing shot

For a beat with no dialogue, just use three beats — opening, action, closing — don't force a dialogue beat that doesn't exist.

The key rule: don't write exact seconds, and never write milliseconds. This one is counterintuitive — you might assume a precise timeline makes the model more obedient, but testing shows the opposite: hardcoded seconds make motion stiffer and actually lower how well the shot executes. The four beats don't need any time markers at all; the model distributes the duration across them on its own.

The only exception is a beat that needs to land precisely on a music cue or lip-synced dialogue. There, use a loose "approximately" range:

Time beat reference:
- Opening: ~0-4s, slow pan across the table
- Dialogue beat: ~4-14s, close up on face
- Final closing beat: ~14-18s, pull back wide

Use a tilde for "roughly N seconds" — never a precise 0.00s–4.00s window.

A full example: every rule in one prompt

Whole scene: A quiet afternoon in an old traditional courtyard house, sunlight slanting across the yard, the scent of osmanthus drifting through the air, shot on 16mm Kodak Portra 400 film, warm soft golden hour lighting, natural film grain, gentle lens bokeh, quiet nostalgic daily atmosphere, no harsh shadows.

Opening shot: Static close up shot of an old wooden chair in dappled sunlight, camera holds still for a beat.

First action beat: Slow pan left, an elderly grandfather @Grandpa sits in the bamboo chair, eyes closed, enjoying the breeze.

Second dialogue beat: Grandpa opens his eyes and smiles, reaching out to gently pat his granddaughter's head.

Final closing beat: Camera slowly pulls back wide, revealing the whole courtyard bathed in golden light.

Reference mapping (strict constraints, no changes allowed throughout the clip):
- Grandpa face, white hair, linen shirt, facial features 100% preserve @Grandpa
- Courtyard layout and architecture strictly match @image2
- Wooden chair and props strictly match @image3

Camera: Very slow gentle handheld long take, almost no camera shake, shallow depth of field, smooth slow focus pull between subjects, no fast cuts, no sudden camera moves.

Sound design: Soft wind through leaves, distant birdsong, no extra background music, only natural quiet ambient sound.

Quality anchor: Photorealistic, natural skin texture, realistic fabric and liquid physics, natural film motion blur, soft contrast, consistent character appearance throughout the whole clip, no floating objects, natural human anatomy.

This example uses every rule covered below: it sets the overall mood up front, keeps the four beats free of timestamps, gives each reference image its own line with a clear responsibility, and closes with the three trailing blocks — Camera, Sound design, and Quality anchor.

Reference images: assign once at the top, then never mention them again

This is the section people get wrong most often, and it's the single biggest swing factor between a clip that holds and one that doesn't.

There's one rule: every reference image's constraint gets written exactly once, in a "Reference mapping" block near the top. After that, whether you're describing action, dialogue, or camera movement, just use the character or object's plain name ("Grandpa," "the osmanthus tree") — don't re-tag it, and don't repeat "preserve the features" language anywhere else.

Why write it this way? The model's attention is a limited budget. Repeating "preserve facial features" in every beat is competing directly with action, timing, and camera language for that budget. State it once up front, clearly, and let the rest of the prompt focus on motion — characters are actually less likely to drift as a result.

The one exception: if you're continuing a shot across a follow-up segment, restate the reference constraint once at the start of the new segment. Within the same segment and scene, continuous action never needs to repeat it — the model carries the earlier setup forward automatically.

The specifics:

1. Give every reference image a name. Clicking a thumbnail inserts an @name tag at your cursor — using a name you chose for that character or object (@Grandpa) helps the model hold onto whose face to preserve far better than a bare @image1. Anything you didn't name specifically falls back to the system's @image1, @image2, in order.
2. Spell out each image's job, one line at a time. Inside the "Reference mapping" block, give each image its own line stating exactly what it's responsible for and how strictly: "100% preserve facial features, hair, and outfit" is a different strength of constraint than "strictly match the scene layout" — write them as two distinct lines, not one blended sentence.
3. Past 3 reference images, always break them into their own block. Don't fold the constraints into narrative sentences — the model can't reliably tell which sentence maps to which image.
4. The cap is 30 images. As long as each one's responsibility is spelled out clearly, a high count doesn't cause them to conflict with each other.

Picking a duration: don't push the ceiling

Not every duration holds up equally well. Testing sorts roughly into three tiers:

If a shot genuinely needs 20+ seconds and involves scene changes or multi-character dialogue, don't force it into a single 30-second generation. Split it into two 12–15 second segments generated separately, then stitch them together in an editing tool afterward — overall stability and motion smoothness both come out noticeably better this way. Packing too many events into one generation raises the odds the model drops something.

Skip negative prompts — say what you want instead

This is the most counterintuitive section, and the one most people get wrong.

You might be used to writing "not ugly, not deformed, no AI look" to steer away from flaws. On Seedance 2.5, this backfires and makes flaws more common. The model has no negative-token filter — "don't do X" gets read as positive description and folded straight into context, where it competes with the weight that should be going to your subject, action, and reference features. The result is that it can conjure up exactly the thing you were trying to avoid.

The fix is simple: rewrite every flaw you want to exclude as a statement of what you want instead. Drop "not ugly, not deformed" entirely and replace it with this positive quality anchor at the end of your prompt:

Quality anchor: Photorealistic, natural skin texture, realistic fabric and liquid physics, natural film motion blur, soft contrast, consistent character appearance throughout the whole clip, no floating objects, natural human anatomy.

(Yes, that line still has one no floating objects in it — that's the one case where a negative phrase is fine, covered in the table below.)

There is one kind of negative statement that should be written as a negative: not a description of a visual flaw, but an explicit instruction telling the model not to do something it would otherwise do by default. These are two very different things:

A hard constraint targets a default behavior the model would otherwise take on its own (like auto-adding a music track), and you have to explicitly call it off. A visual flaw is a random-error category — describing what you want positively works far better than listing what you don't want.

A cheat sheet: camera, film, and lens terms

You don't need to invent phrasing — these are the terms that tested with the highest fidelity. Copy them directly.

Camera movement

Not recommended: whip pans, fast cuts, bullet time, extreme rapid motion — these categories test noticeably lower and are prone to warping or freezing. Keep the number of camera beats in a single generation to 2–5, and give each beat at least 3 seconds to hold. If you genuinely need 0.8–1.5 second fast cuts (trailer-style), generate each shot as its own 4-second clip and do the fast cutting in an editing tool afterward — don't ask the model to execute rapid cuts inside a single generation.

Film stock and texture

Lens length (write "focal length + lens type + shot size" together — it beats vague depth-of-field description on its own)

For a big emotional facial close-up, use 85mm f/1.4 portrait lens, extreme close up shot; for hand movements or small props, use 100mm macro prime lens, close up shot — these two combinations test with the highest fidelity in their respective use cases. Example:

Camera: 85mm f/1.4 portrait prime lens, extreme close up on character face, slow subtle breath movement, shallow depth of field, soft warm side lighting.

Which genres are naturally more stable

Success rates aren't uniform across genres — worth knowing going in:

• Very stable, safe to experiment freely: neo-noir, 1990s street realism, nostalgic everyday scenes, quiet slice-of-life, crime drama, art-house film.
• Mostly stable, a few tries usually gets there: light fantasy, martial-arts action, urban night scenes, near-future sci-fi.
• Not recommended for regular use: pure cartoon animation, extreme VFX blockbuster style, surreal abstraction — these genres just aren't what this photorealistic-leaning format is built for.

If you're shooting sci-fi or mecha subjects, leaning harder on the photorealistic film terms and restrained camera vocabulary above helps pull overall stability up.

A last pass before you hit generate

Run through this checklist on any prompt before submitting it:

• Is camera/lighting/texture/action in English, and dialogue/on-screen text fully in your target output language?
• Did any of the four beats accidentally pick up exact seconds? (Fine only for a music- or lip-sync-cued beat.)
• Are reference constraints written once at the top, with plain names used everywhere after that?
• Is a complex shot over 20 seconds, with scene changes or multi-character dialogue, split into two segments instead of forced into one?
• Any leftover "not ugly, not deformed" phrasing? Rewritten into a positive Quality anchor?
• Does the prompt close with all three trailing blocks — Camera, Sound design, and Quality anchor?

Run a prompt through this list once, and it'll read like the examples above.