Kling 3.0（含 3.0 Turbo）｜提示詞教學（繁體中文）

1. 一句話定位

Kling 首個把影片、音訊、圖像統一在同一架構的版本。3–15 秒、720p/1080p/4K、16:9/9:16/1:1、音訊可開關、最多 5 個鏡頭（各有提示與長度）、首／尾幀、8 個預設。 官方型錄定位為「較低成本、單一平面、不需要重動態」的選擇；Turbo 為快速簡單動態版。

官方定位（原文）：Lower-cost option for single-plane scenes that don't need heavy motion. Multi-shot, audio sync, motion transfer.
— 來源：skills-catalog

官方定位（原文）：Fast Kling option for simple motion. Text-to-video and single start-frame animation when the user explicitly wants speed, lower cost, or a quick Kling 3.0 variant.
— 來源：skills-catalog

2. 規格速查

3. 提示詞結構公式

[Subject] + [Action] + [Camera behavior] + [Visual style]
Multi-shot (custom): Shot n = one action + one camera move; tag @element in every shot it appears

官方步驟：「describe the subject, the action, the camera behavior, and the visual style」。多鏡頭時每鏡只承載一個動作與一個鏡頭運動；@element 的外觀已由元素定義，所以鏡頭提示只描述它的動作。

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

flowchart TD
  A[總長 3–15s] --> B{多鏡頭?}
  B -->|草稿| AU[Auto：模型自行分鏡]
  B -->|精確| CU[Custom：每鏡 1 動作 + 1 鏡頭運動]
  CU --> E[每鏡都 tag @element]
  AU --> F{首/尾幀?}
  E --> F
  F -->|有| T[提示只寫兩幀之間的過渡]
  F -->|無| G[直接生成]
  T --> D[720p 草稿 → 定稿再升 1080p/4K]
  G --> D

4. 逐步教學流程

1. 在 Video 選 Kling 3.0，決定長度與比例。
2. 寫提示：主體、動作、鏡頭行為、視覺風格。
3. 需要時開多鏡頭：Auto 適合草稿；Custom 逐鏡指定提示與長度。
4. 需要一致的角色／產品，用 @elements 並在每個出現的鏡頭都 tag。
5. 有首／尾幀時，提示只描述兩幀之間發生甚麼及鏡頭如何移動。
6. 以 720p 起草，鎖定提示與結構後才以 1080p／4K 定稿。

5. 技巧與範例提示詞

5.1 @character 元素＋時間碼分段（15 秒牛仔競技）

範例提示詞（英文原文，逐字引用）：

Cinematic live-action rodeo film, 15 seconds, golden-hour sunlight, starring @character, the blonde cowgirl from the reference: long voluminous honey-blonde waves, cream felt cowboy hat, white crop top, baggy blue jeans with brown belt, brown embroidered western boots. 100% matches the reference in every frame.

0-1.5s: inside a wooden bucking chute, backlit by low sun, @character sits astride a massive black rodeo bull, right hand pressing her hat down; a handler's arm pulls the gate rope; dust hangs in the warm rim light. At 1.5s the gate swings open.
1.5-3s: side profile tracking shot: the bull lunges out of the chute into the arena, she rides one-handed with her hat hand raised, hair and fringed chaps whipping; sun flares through the dust, packed grandstand behind.
3-4s: HARD CUT to a tight backlit close-up of her face under the hat brim, calm confident eyes, golden edge light in flying hair; a whip-pan through white fabric transitions out.
4-8s: wide arena coverage shot from behind silhouetted spectators, their hats and raised hands framing the foreground: the bull bucks and spins in the center of a sun-hazed dirt arena, dust clouds blooming with each kick; she stays balanced, torso countering every buck, one hand high. Rodeo safety clowns in green shirts circle at a distance.
8-10s: frontal shot: the bull charges toward camera through the dust, she rides tall, hand on her hat, crowd roaring in the stands.
10-11.5s: she leaps off mid-buck, a dynamic tumbling dismount, dust exploding on landing.
11.5-13s: low ground-level shot from behind her legs: brown boots with spurs planted in the dirt in sharp foreground, the bull walks away into the haze screen-right, a clown waves it off screen-left.
13-15s: she runs toward camera laughing with motion blur, then stops in a medium hero frame: wide smile, one hand touching the hat brim, golden backlight, cheering grandstand behind, the exact reference face.

Physics: real bull mass and gait, dirt kicked in arcs, hair and cloth follow momentum, dust lingers in the air. Lighting: low warm sun, strong backlight and lens flares, faces lifted by bounce off the dirt. Style: anamorphic western cinema, warm Kodak tones, natural film grain, handheld energy on wide shots, crowd ambience and hoof impacts, no subtitles, no on-screen text.

中文解說：官方範例：首行定類型、長度、光線與 @character（並重述參考的關鍵外觀），之後以 0-1.5s、1.5-3s… 時間碼逐段寫動作，「HARD CUT」標明剪接，最後 Physics／Lighting／Style。

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

5.2 首幀控制：從首幀位置開始的連續一鏡

範例提示詞（英文原文，逐字引用）：

0.0-1.5s: Following his glance, he lifts his arm off the wheel, swings his legs out and stands up beside the open door, squinting into the light.
1.5-3.0s: A man in a rumpled band uniform sprints up the driveway out of breath and PRESSES a battered brass trumpet into his hands, gesturing toward the lawn; a dozen more musicians hurry in with tubas, trombones and a bass drum, forming a ragged semicircle around the car, watching him expectantly.
3.0-6.0s: He looks at the trumpet, at them, then lifts it, sets the mouthpiece, cheeks filling, and BLOWS a huge bright opening note, shoulders rising with the breath, fingers dropping onto the valves. On his second phrase the whole band CRASHES in behind him, drum kicking, tubas thumping, the group stepping into rhythm.
6.0-8.0s: He walks two steps forward still playing, the band falling in around him, faces appearing at the mansion windows; the frame holds him mid-phrase, eyes closed, fully committed, as the shot ends.

CAMERA: starts exactly at the start-frame position: outside the car, three-quarter front at chest height, framing him through the open door, loose handheld, drifting slightly right as the band assembles, then holding on him. No cut.
PHYSICS: real trumpet technique: correct embouchure, visible breath support, valves moving with the notes; brass has real mass and swings on straps; the drum head flexes on each hit; the door swings naturally as he stands.
LOOK: continuous with the start frame: bright clear daylight, balanced exposure, detail kept in the white facade, no blown-out highlights, neutral white balance, matte skin, the pink shirt and steel-gray car keep their exact start-frame colors, 35mm film grain, no CGI smoothness. No logos, no readable text.
AUDIO: quiet neighborhood ambience, birdsong, hurried footsteps, one breathless shout, the single bright trumpet note, then the full brass band swelling in: live and slightly rough.

中文解說：官方範例：CAMERA 段寫「starts exactly at the start-frame position」；LOOK 段要求與首幀顏色連續（粉紅恤衫與鋼灰色車保持首幀顏色）。

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

5.3 一鏡到底＋POSITIVE LOCKS（6 秒更衣室）

範例提示詞（英文原文，逐字引用）：

SCENE CONTEXT Six-second one-take: a red-haired girl in a navy tracksuit on the floor of a deep-red locker room: jacket-fix, room-scan, then a level stare into the lens as the handheld camera pushes in to a close portrait.

ACTIVE REFERENCES @GIRL: 100% per reference: copper-red hair pulled back, loose strands; steel-blue eyes; navy track jacket, white chest panel, navy pants, dark-red sneakers. Seated, one knee up. Contained defiance: fix as armor-adjustment, scan as a fighter reading exits, the stare a statement. @LOC: per reference: deep-RED locker room: crimson panels, red lockers with chrome latches, white tile, one cool overhead pool into red gloom.

FORMAT / CAMERA ONE UNBROKEN TAKE, 6.0s, zero cuts. First frame = reference composition, already alive: hands on the hem, eyes down, no empty start. 47 degrees natural, rectilinear, focus locked on her face. MAXIMUM HANDHELD: breath-sway, micro-corrections, framing by feet, never zoom, never stabilized. ONE continuous move: slow walking push 2.5m to 1.2m, waist-up to close portrait, at its closest exactly as her eyes land on the lens.

ACTION TIMING 0-2.5s, THE FIX. Head down. Both hands tug the hem, squaring the white panel; collar flick, a copper strand falls across her temple; she leaves it. The zipper stays untouched, fingers never near it. Camera breathes, then advances on the collar flick. 2.5-4.5s, THE SCAN. Hands settle on the knee. Eyes sweep LEFT into the red depth, hold, track RIGHT along the chrome latches: reading the room, not the lens. Camera closing. 4.5-6.0s, THE LOOK. Eyes come off the lockers, land level on the lens. Chin lifts a degree. One slow blink, then stillness, breath in the shoulders. Camera settles into close portrait, holds. Ends MID-STARE.

PHYSICS / LIGHT / AUDIO Real breath, true blink timing, nylon rustle, the strand staying. One cool overhead key carves cheekbones and the white panel; red gloom, red bounce, chrome speculars; her face takes more key as the camera nears. Ventilation hum, fabric rustle, one nose-exhale. No music, no dialogue.

POSITIVE LOCKS Order locked: fix (zipper untouched), scan left-right, eyes land on the lens and HOLD. Eyes meet the lens exactly ONCE, at the end. ONE camera move. ONE person; 100% per reference. 8K photoreal, 180 degree shutter, pore-level skin, true fabric/hair physics; no 3D render, no plastic skin. NO IP / NO BRANDS / NO LOGO

中文解說：官方範例：以 POSITIVE LOCKS 列出必須保持的次序與限制（拉鏈不碰、眼神只對鏡頭一次、只有一個鏡頭運動、只有一人）。用正面「鎖定」代替負面詞。

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

5.4 微距一鏡＋聲音（OPTICS、FIRST FRAME 座標）

範例提示詞（英文原文，逐字引用）：

SCENE CONTEXT A single continuous extreme macro push-in on a scratched vintage military field watch held in a man's fingers, from a full view of the dial into an extreme close-up of the aged hands at its center.

FIRST FRAME AND BLOCKING Frame one shows the full watch filling 80% of frame width, dial at x 55%, y 50%, tilted slightly toward camera against a near-black void with faint warm falloff. Weathered fingers grip the case at frame left and lower right, out of focus. Crown points screen-right. Time reads 7:37, second hand near 6.

Watch: unbranded 1970s military field watch. Matte black dial speckled with dust and micro-scratches under a domed crystal. Cream aged-lume Arabic numerals 1-12, railroad minute track, small broad-arrow mark below center. Cathedral hour hand with two rounded lume lobes, pencil minute hand, thin steel second hand. Brushed steel case, worn chipped bezel, knurled crown.

FORMAT Single continuous take, no cuts, no speed ramps.

OPTICS Macro probe-lens character, 18 degree field of view narrowing into true macro. Camera starts 25 cm from the dial, ends 3 cm from the hands. Razor-thin focus tracks forward, keeping the hands sharp while numerals and case melt into warm bokeh; the final frame resolves lume grain and dust fibers.

CAMERA Slow hypnotic dolly-in along the lens axis toward dial center, constant velocity, no drift, no reframing; precision-slider smoothness, not handheld. The hands' crossing point settles just right of center, flanked by the 8 numeral and the broad arrow.

ACTION TIMING 0.0-2.0s: full dial in frame, fingers gripping; second hand ticks; push-in begins. 2.0-4.5s: fingers exit frame edges as magnification grows; numerals slide past; focus tracks the hands. 4.5-6.0s: extreme macro on the layered hands over the black dial; dust and scratches fully resolved; second hand keeps moving to the last frame.

PHYSICS Second hand advances with real mechanical cadence. Faint micro-tremor from the holding hand early on. Dust motes drift through the light near the crystal. No morphing of numerals or hands.

LIGHTING Warm tungsten-amber key from upper left, raking low so every scratch and embossed numeral casts a tiny shadow. Deep low-key: background crushed near-black, highlights only on case rim, hand edges and lume. Golden patina palette, no flat front light.

AUDIO Soft close-mic ticking, faint room tone. No music.

LOCKS No logos or readable text besides numerals and minute track. Photoreal macro, fine film grain.

中文解說：官方音訊段範例：FIRST FRAME 用百分比座標定位（錶盤佔畫面寬 80%，x 55%、y 50%），OPTICS 寫出起止距離（25cm → 3cm），AUDIO 只要近收音滴答聲、無音樂。

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

5.5 CLI：首幀＋pro 模式

範例提示詞（英文原文，逐字引用）：

higgsfield generate create kling3_0 \
  --prompt "slow camera push through a forest clearing at dawn" \
  --start-image ./first.png \
  --duration 5 --mode pro --sound off \
  --wait

中文解說：官方 README 範例（譯：黎明時鏡頭緩慢推入森林空地）。

來源：cli-readme

6. Do / Don't 清單

來源：blog:Kling-3.0-is-on-Higgsfield-User-Guide-AI-Video-Generation

7. CLI 參數表（自動擷取自官方 MODELS.md）

自動解析 cli-models；以 higgsfield model get <id> 取得即時 schema。

kling3_0 — Kling v3.0

kling3_0_turbo — Kling 3.0 Turbo

higgsfield generate create kling3_0_turbo --prompt "fast handheld product reveal on a clean studio table" --start-image ./first.png --duration 5 --resolution 720p --wait

來源

詳見同資料夾 sources.md。所有事實存取日期：2026-10-08。