Cinema Studio Pro-Skill

Skill is the recipe card system inside the Studio's video mode.

You write what to shoot. The recipe handles how — when you hit Shoot, your one-liner gets expanded into a full production prompt with the shot sizes, camera positions, movement, action physics, sound layers and continuity rules that the genre actually needs.

Same sentence, "two people fighting in a rainy alley," with and without a card: two completely different films.

Three things to know first

1. The recipes are tuned for Seedance 2.5.

The camera language, stage structure, reference-binding syntax and sound notation baked into each card are all written to Seedance 2.5's spec. So when you pick a genre card while on Seedance 2.0, the Studio switches your model to Seedance 2.5 for you and shows an "Applied · …" toast telling you exactly what changed.

That isn't the app overriding your choice for no reason. The two models pull in opposite directions: 2.0's refine pass exists to make prompts tighter, while a recipe demands every stage be written out in full — picture and sound. Stack them and the model gets two contradictory instructions, and quality quietly becomes a coin flip.

If you're already on Seedance 2.5, picking a card won't touch your model. And two cards — AI Refine and Idol Stage — work on any model and never change your selection.

2. Skill only exists in video mode.

There's no Skill button in image mode (Develop). There's none in video editing, video extension, or lip-sync mode either — in those, your prompt isn't the main input.

3. Skill is a paid feature.

Free plans see every card, but the button carries a lock; clicking it opens the upgrade notice.

A note on card names

The recipe cards are currently labelled in Traditional Chinese in the picker. Each section below leads with an English name and gives the on-screen label right after it, so you can match them up:

Kung Fu Cinema — 功夫武打電影

Pick by the icon and the Chinese label; read this page for what the card actually does.

Three steps to use a card

1. Open the picker — Studio → switch to Videos → the Skill button (✨) in the toolbar under the prompt box. The panel is titled "Pick a Skill recipe" and has a search field at the top.

2. Pick a card. Three things happen:

• A chip with an icon appears at the very start of your prompt box. That chip is the recipe.
• If needed, the model / aspect ratio / duration get set for you, with an "Applied · …" toast spelling out what changed.
• If the card says something like "Suggested: attach 2 reference images," go attach them.

3. Write what you're shooting, then hit Shoot. Type after the chip as normal. You don't need to mention your uploads in the prompt — attaching them is enough. To point at a specific one, click its thumbnail in the reference slot and a token (@圖片1, @音訊) gets inserted at your cursor.

Behaviours you will definitely hit

The two cards that aren't about genre

AI Refine — AI 優化

Use it when you don't know which card fits, or what you're shooting isn't one of the genres below. It runs the standard refine pass your current model would run anyway, and it's the card mounted by default.

Write one clear sentence about the picture: "outside a convenience store at midnight, rain on the glass door, a black cat sitting by the entrance licking its paw."

Note — the only card whose behaviour is identical on every model. When in doubt, this one is never wrong.

Port to This Model's Format — 轉成本模型格式

Use it when you already have a finished prompt — carried over from Seedance 2.0, copied from another platform, handed to you by someone else — and you want it reformatted for the model you're on.

Write nothing new. Paste the whole thing in.

What it does — changes the writing, not the story. Character count, event order, location, dialogue and ending state are all preserved. It redistributes scattered information into the official structure and strips markup from other models (leave that in and it gets read as picture content).

Note — this is the only task card; every other card is a genre card. It doesn't decide what you shoot, only how it's written. That also makes it the best bridge for 2.0 users: pick it on 2.0 and it moves you to 2.5 and converts your prompt in place.

Action and fights

High-Speed Vehicle Chase — 高速載具追逐

Use it for full-speed pursuits with cars, bikes or boats, including loss of control, rolls, collisions, fire and evasion.

Attach 2 reference images (character + vehicle, or one per vehicle).

Write who's chasing whom, on what road surface, and whether you want a crash: "a black superbike chases a silver sedan across a rain-slicked overpass; the sedan loses control and hits the barrier."

Note — this card lives on physics: grip, vehicle weight, suspension, debris direction, the source of the explosion, the rider's control inputs and the recovery. It deliberately refuses unmotivated explosions, floaty rolls and crashes where the wreck resets between shots. Lots of events — don't set the duration too short.

Modern Close Combat — 現代近身格鬥

Use it for grounded contemporary fighting with no martial-arts school attached: close-quarters exchanges, real impact, spatial advantage, disarms. Spies, alley brawls, corridor fights.

Attach 2 reference images (one per fighter).

Write who's against whom, where, and who has the upper hand: "an operative pinned in a narrow stairwell by two attackers, takes the knife off one of them."

Note — every exchange is written out end to end (stance → attack path → point of contact → impact → displacement → recovery), and your protagonist will get hurt, lose ground, and then turn it around. If you want an untouchable hero, this isn't the card. For named martial-arts styles, use the next one.

Kung Fu Cinema — 功夫武打電影

Use it for Hong Kong-style martial arts: the stillness before the strike, a cut to the hand as it goes out, a micro-detail on contact, the camera thrown by the impact, the settle at the end. Pick this when you want cinematic fighting rather than just hard fighting.

Attach 2 reference images (one per fighter).

Write the style in — this is the card's biggest lever: "two women fight in a rainy alley, one trained in Wing Chun, the other a street fighter."

Styles the card knows: Wing Chun, Tai Chi, Hung Gar, Muay Thai, boxing, karate, taekwondo, judo/jiu-jitsu, drunken boxing, praying mantis, bagua, xingyi, street fighting, plus weapons (sword / blade / staff). Name a style and the techniques come from that style — and when two styles meet, the card writes the clash between them. Wing Chun's centre-line chain punches driving someone back into a wall, the street fighter grabbing a chair to break that short-range rhythm: that collision is why Hong Kong fight scenes work. Name nothing and you get generic southern-style hands and feet.

Three things to know:

• Say so if you want flying. Ground-based by default. Write "wuxia" or "wire work" to unlock leaping, wall-running and mid-air exchanges.
• Two takes looking different is intentional. Anything you don't specify — location, weather and light, opening camera position, dominant camera move, contact detail, who strikes first, rhythm, ending — gets redrawn on every submission, so back-to-back takes don't come out identical. Pin any of them down by writing it in your prompt; whatever you specify always wins.
• Don't mix styles. "Wing Chun high kick" or "Tai Chi haymaker" defeats the card's technique dictionary.

Music and performance

Idol Stage — 偶像舞台

Use it for K-pop-style group stage performance: choreography, formations, follow spots, a sea of lightsticks.

Attach 1 reference image (member look).

Write the formation and the beat: "five-member girl group on a circular stage, synchronised choreography, V-formation push forward as the chorus drops."

Note — this card avoids slow motion by default (overcranking kills the sense of rhythm) unless you ask for it. It treats lighting as the lead, and it writes expression and eyeline, which is what makes an idol shot land. One of the few cards that won't switch your model.

Beat-Synced Music Video — 節拍同步 MV

Use it for music videos where the cutting is on the beat: multi-shot, formation handoffs, lead-dancer coverage, a final downbeat freeze.

Attach 1 reference image + an audio reference.

Write the BPM, who leads, and which beat you cut on: "four-member group, 128 BPM, cut on the chorus downbeats, eight bars of solo for the centre dancer."

Note — this card is dance plus beat-cutting, not singing. For characters singing along, use the next one.

Stylised Singing — 風格化演唱

Use it for uploading a song and having your character lip-sync to it. The mouth follows the vocal, but what you get is a stylised interpretation — Seedance 2.5 reworks the melody and the delivery on its own.

Required: an audio reference (required). One reference image is also recommended.

⚠️ Without audio the whole submission is blocked — and you are not charged. You'll be warned when you pick the card, and again if you press Shoot. Lip-sync with no audio has nothing to sync to; generating anyway would just get you a song the model invented.

Write who's singing and where: "a singer at the studio microphone wearing monitor headphones."

Note — the audio you upload is the song, and the submission does tell the model to stay faithful to it — no rearranging, no added lyrics. In practice Seedance 2.5 still reworks the song and the performance: expect something closer to a cover version than a word-for-word reproduction. That's where the model sits today, not something you got wrong. What the card does handle: mouth closed through intros, instrumental breaks and gaps between lines; breaths written into shoulder and chest movement; at least one shot where the mouth is clearly readable; room acoustics matched to the location. Full-length slow motion is forbidden — overcrank the footage and the sync is gone for good.

Story and character

Emotional Oner — 情緒一鏡到底

Use it for an unbroken single take carrying an emotional arc. Someone sits still inside chaos, gets up, walks through connected spaces, stops somewhere quiet — that shape.

Attach 1 reference image (the lead).

Write the emotional starting point, the spaces to cross, and the small action at the end: "she sits with her head down through a loud wedding banquet, gets up, walks the corridor to the back door, leans on the wall and lights a cigarette."

Note — the spaces have to actually connect. Cuts, teleports, unmotivated orbits and spatial breaks are all forbidden, and the sound falls away as the character moves.

Natural Two-Hander Dialogue — 自然雙人對話

Use it for two people talking, sitting or standing, where the goal is lived-in rather than interviewed.

Attach 2 reference images (one per character).

Write your dialogue out verbatim with speakers marked, then add relationship and setting: "A: How've you been? B: Same as ever. There are layoffs coming." — "two old university friends at a late-night barbecue stall."

Note — the card locks the left/right positions and the axis, assigns every line to a speaker, and keeps the listener's mouth shut, reacting with the eyes (two people mouthing along simultaneously is the classic AI-video disaster). Long dialogue in a short duration will overrun, so budget it.

Documentary and everyday life

Solo Selfie Vlog — 單人自拍 Vlog

Use it for one person filming their own day across several clips.

Attach 1 reference image (the lead).

Write the theme of the day and the places: "a solo trip through Tokyo — breakfast counter, subway, a bookshop, an izakaya at night."

Note — this card is really about who is holding the camera. It allows exactly two answers: the subject holds it (written out with arm tension, walking bounce, off-centre framing) or it's propped on a specific surface (height, distance, tilt, and how the subject walks over to retrieve it). Invisible camera operators, unmanned tracking, drones and third-person orbits are all forbidden — the moment one shows up it stops reading as a selfie.

Reality-Show Test — 真人秀測試

Use it for TV-style destruction testing where the evidence has to be legible. Durability tests, product limits, prop testing.

Attach 2 reference images (tester, object).

Write what's tested, how, and how many rounds: "a smith tests a blade he forged — rope, then bamboo, then steel pipe; the edge chips on the third round."

Note — slow motion is used only on the moment of contact, as evidence, never across the whole scene. Damage accumulates into the next round instead of resetting. Clean HD television look, not film-stock grading.

Wildlife Documentary — 野生動物紀錄片

Use it for long-lens observation of real animal behaviour and habitat, narration optional.

Attach nothing — just name the species clearly.

Write species, relationships, time of day and weather: "misty dawn on the savannah, a cheetah mother leads three cubs toward the waterhole."

Note — the animals stay animals: real micro-behaviours, natural physics, native sound. Pet-like performance, cartoon eyes, anthropomorphic smiles and impossibly close lenses are excluded. Narration explains survival and ecology rather than describing what you can already see.

Spectacle and rendering style

Live-Action Spectacle — 真人電影奇觀

Use it for enormous practical-feeling set pieces: disaster, collapse, demolition, extreme stunts.

Attach nothing.

Write the scale contrast and the decisive moment: "a lone figure at the top of a failing dam, the whole wall coming down behind them."

Note — this card runs on an intensity curve, not a stunt count: silence → ignition → skill on display → spatial breach → sensory peak → sudden silence → decisive image → pull back to reveal. Every explosion, water column and piece of debris needs a source. If you want wall-to-wall intensity this is the wrong card — its power comes from that sudden quiet.

2D Animated Action — 2D 作畫動作

Use it for cel-shaded anime-style fights. Not live action, not 3D.

Attach 1 reference image (character design).

Write the design, weapons and energy colours: "a white-haired swordsman and a red-armoured warrior clash atop a collapsing tower; the blade energy is cyan."

Note — the medium is locked to finished cel-shaded 2D animation: it won't drift into live action or real skin, and no rough construction lines appear. It writes to animation beats (anticipation key → smear on the acceleration → clean contact key → sparks and cloth follow-through → settle into silhouette), which is why the action stays readable instead of turning to mush.

From previz to final

Previz-to-Final Conversion — 多參考預演轉換

Use it for turning a blocking pass, previz or rough clip you already have into finished footage, keeping its cutting and framing.

Attach a video reference plus 2 or more reference images (colour keyframes, depth passes, characters, props).

Write what each piece of material is responsible for, and how many people are in frame: "@圖片1 is the lead's appearance, @圖片2 is the spatial depth relationship, there is only one person in the entire shot."

Note — the most advanced card here; the previz's cutting, framing and timing are hard constraints. What it cares about most is how many people exist: whether the front, the back and that hand in the foreground are the same person. Leave it ambiguous and you grow a second person, a shadow person or a mirror person. With fewer than two reference images, another card will serve you better.

Quick reference

FAQ

I picked a card and my prompt didn't change at all. Is it working?
Yes. The recipe expands in the cloud the moment you hit Shoot, and the expanded version never comes back to your screen. Your box only ever holds your own words — so you never have to wonder what you actually submitted.

Can I attach two cards at once?
No. One card is one complete recipe; picking a second replaces the first. That's deliberate — two recipes stacked would hand the model contradictory instructions.

Is "Suggested: attach 2 reference images" mandatory?
No, that's a suggestion — you can generate with fewer (the result just won't be as strong). The one hard requirement is the audio reference on Stylised Singing, which blocks the whole submission and doesn't charge you.

Can I write my prompt in English?
Yes. Your dialogue, narration and on-screen text are preserved verbatim in whatever language you wrote them, and the recipe declares the language and accent so the delivery matches.

I chose a different model — why did it switch me to Seedance 2.5?
Because that recipe can't run on the model you were on. You get an "Applied · …" toast naming what changed, and anything you adjust afterwards sticks. Switch back and, if that card isn't compatible, the chip removes itself and tells you.

Why is there no Skill button in video editing / extension / lip-sync mode?
In those modes the main input isn't a prompt — it's the source clip, the last frame, the vocal track. A recipe would have nothing to act on, so the button simply doesn't appear.

Two Kung Fu takes came out completely different. Is that a bug?
No, it's the design. Anything you don't specify gets redrawn on every submission — location, weather and light, opening camera position, dominant camera move, contact detail, who strikes first, rhythm, ending — so back-to-back takes don't collapse into the same film. Pin any of it down by writing it in your prompt.

Do I get charged if the expansion fails?
No. If the recipe fails to expand or comes back empty, the whole submission stops there and tells you nothing was charged. It won't quietly fall back to sending your raw text — that would mean paying for a recipe and getting a film that looks like you never picked one.

Is there a limit on how many times recipes can expand?
There's a daily AI text quota tied to your plan, in the hundreds per day on paid plans — normal creative work won't reach it. If you do hit it you'll be told directly, and again, no charge. Generations without a card attached don't consume it.

Can I use this on the free plan?
You'll see every card, but the button carries a lock and clicking it opens the upgrade notice. Skill is a paid feature.

New to the Studio itself (projects, the asset wall, reuse, editing an image by describing the change)? Start with the Studio Handbook. Want a full short film with a story and consistent characters? See the Short Film Handbook.