跳到正文
DEV Community· SetPosBoy·· 13 小時前AI 評分59

如何為 AI 影片模型撰寫提示詞:六部分公式(Seedance、Kling、Veo)

How to write prompts for AI video models: a six-part formula (Seedance, Kling, Veo)

AI 導讀

文章提出一套適用於 Seedance 2.5、Kling 3.0 和 Veo 3.1 等 AI 影片模型的六部分提示詞公式,涵蓋主體、動作、場景、鏡頭、光線與風格,以及聲音。提示詞應先寫主體和動作,並把風格放在末尾;圖生影片則只需描述動作、鏡頭和聲音。測試時可固定 seed、每次只改一部分並記錄結果;如服務支援,可先以最低解像度和最短片長試跑。

正文

Video models like Seedance 2.5, Kling 3.0 and Veo 3.1 understand plain text, but a short prompt such as "a girl dances on the street" almost always gives a random result: the camera jitters, the motion is wrong, and the lighting changes on every run. Below is a six-part formula that makes generation predictable, a way to test prompts without burning credits, and three before/after examples. The examples use Seedance 2.5, but the approach carries over to other models.

How a video prompt differs from an image prompt

In an image, the model decides what is in the frame. In a video it has to decide three more things: what changes over time, how the camera moves, and what you hear, if the model generates sound. Whatever the prompt leaves out, the model fills in on its own, and differently each time. That is where drifting faces, extra movements and random cuts come from.

So a good video prompt describes not only the frame but the action over time.

The six-part formula

Part What to write Example
1. Subject One main subject and 2–3 details a young woman in an oversized white hoodie
2. Action One action for the whole clip dances to an upbeat beat
3. Place Location and time of day a neon-lit Tokyo street at night
4. Camera One camera move per clip slow orbit around her
5. Light and style Lighting, texture, genre rain reflections, cinematic light
6. Sound If the model generates audio the beat and city traffic

Subject. The fewer characters, the more stable the result. The details help the model keep the same look from the first frame to the last.

Action. "Dances to an upbeat beat" works. "Dances, then sits down and drinks coffee" does not fit into 5 seconds: the model either drops part of it or blends the actions together.

Place. The location sets the background and half of the lighting, so always include it.

Camera. Static shot, slow push-in, orbit, handheld, side tracking. Two camera moves in one short clip usually merge into one vague motion.

Light and style. "Film", "documentary", "commercial" change the result a lot even with the same story.

Sound. If the model does not generate audio, skip this part.

Order matters

Start with the subject and the action, and put the style at the end. In practice models lean on the beginning of the prompt: if it opens with "cinematic, 4K, film" and the hero appears at the end, you get the style but not the motion.

Camera vocabulary

  • static shot: the camera does not move, good for product shots;
  • slow dolly in / dolly out: a smooth push-in or pull-out;
  • orbit around the subject: the camera circles the hero;
  • tracking shot: the camera travels next to a moving subject;
  • handheld: slight shake;
  • low angle / high angle: shot from below or above.

When the first frame is an image

In image-to-video mode the look is already in the picture. Do not describe it again, or the model starts "fixing" the face and clothes. Write only the motion, the camera and the sound: "she turns her head and smiles, slow dolly in, soft city noise".

How to test a prompt without burning credits

A video generation costs much more than an image, so debug a prompt the way you debug code: in small steps, writing down results.

  1. Draft first. If the service allows it, make the first attempts at the lowest resolution and shortest duration. Composition and motion are already visible there, and the final version runs once.
  2. Fix the seed. If the model supports a seed, fix it while you debug. Then differences between runs come from your prompt changes, not from chance.
  3. Change one part at a time. Change camera and light together and you no longer know what worked.
  4. Keep a log. A simple table: prompt, what changed, what happened. After a dozen runs you have your own list of phrasings that work on a given model.

Credits matter here because the same clip costs very different amounts depending on the plan. We priced one Seedance 2.5 clip (720p, 5 seconds) on six public plans on Nov 20, 2026:

Price of one Seedance 2.5 clip on six plans, Nov 20, 2026

Plan Price per clip
POINT Creator $1.04
Higgsfield Ultra $1.18
POINT Pro $1.42
Higgsfield Plus $1.71
Syntx Elite $2.44
Syntx Basic $3.63

With ten test runs per final clip, the difference between the cheapest and the most expensive plan is about $26 per finished video.

Three examples

1. Vertical dance clip, 9:16. Before: "a girl dancing on the street at night". After: "A young woman in an oversized white hoodie dances to an upbeat beat on a neon-lit Tokyo street at night, camera slowly orbits around her, rain reflections on the asphalt, cinematic, sound of the beat and city traffic". All six parts, one action, one camera move.

2. Product shot. Before: "beautiful sneaker ad". After: "A white sneaker slowly rotates on a wet black stone, static shot, soft studio light from the left, water drops on the surface, premium commercial style". For a product, keep the camera still: the object provides the motion.

3. Background landscape. Before: "mountains, sunset". After: "Fog rolls over a mountain ridge at sunset, slow dolly out, warm golden light, birds in the distance, calm ambient wind". Here nature does the action: the fog moves, the camera pulls back.

What to avoid

  • Text in the frame: models still often distort lettering and logos.
  • Close-ups of hands and fast, complex finger movements.
  • Several actions in a row in a short clip.
  • Conflicting styles: "anime, photorealism, film" in one prompt.
  • Long lists of adjectives without an action.

Checklist before you generate

  1. Is there one main subject with 2–3 details?
  2. Is there one action for the whole clip?
  3. Are the place and time of day set?
  4. Is there one camera move?
  5. Are style and light at the end?
  6. No text in the frame and no extra characters?

A couple of minutes on the prompt usually saves several regenerations.

Turn a photo into a video in three steps

Disclosure: I build POINT, an AI studio with 100+ models (Seedance 2.5, Kling 3.0, Veo 3.1, Wan 3.0, Nano Banana, Suno, ElevenLabs) on one balance, where I tested these prompts. The formula works the same in any service.

來源:DEV Community · dev.to