# Seedance 2.5 Prompt Guide: Structure, Long Takes and 6 Prompts

Seedance 2.5 is a video model from ByteDance, and it is the default model in the Seedance25 [generator](https://seedance25.com/#generator). It renders clips from 4 up to 30 seconds, at resolutions from 480p to 4K, and it can generate native audio alongside the picture. A model with that much range rewards a clear brief. This guide walks through a prompt structure you can reuse, how to plan long takes, how to get the most out of reference images and first and last frames, when to switch audio off, and six prompts you can paste in today.

Seedance25 is an independent service and is not affiliated with ByteDance. We build the studio around the model and write a lot of prompts for it.

## What a good Seedance 2.5 prompt does

Think of the prompt as a short shot description written for a camera crew. Seedance 2.5 needs to know what is on screen, what changes over time, how the camera behaves and what the scene sounds like. When one of those is missing, the model guesses, and guesses tend to be bland.

A few principles hold up across almost every render:

- **Use full sentences.** Plain descriptive language gives the model relationships between things, which a list of keywords cannot.
- **Lead with what matters most.** The subject and its action belong at the start. Mood words can come later.
- **Be concrete.** "A copper kettle starts to steam on a gas stove" is something the model can show. "A cozy vibe" is not.
- **Keep it consistent.** If you ask for a rainy night, do not also ask for harsh midday sun.

## The five-part prompt structure

This order works well for text-to-video and for the other modes too. You can skip parts, but keeping the sequence makes prompts easy to read and edit.

| Part          | What it covers                              | Example                                             |
| ------------- | ------------------------------------------- | --------------------------------------------------- |
| Subject       | Who or what, with the details that matter   | "A teenage skateboarder in a faded green hoodie"    |
| Action        | What happens, in the order it happens       | "rolls down an empty street and kicks the board up" |
| Camera        | Framing and movement                        | "low tracking shot beside him, then a slow tilt up" |
| Light / style | Time of day, light source, look of the film | "late afternoon sun, warm haze, 16mm film texture"  |
| Sound         | Ambience, music, dialogue, or none          | "wheels on asphalt, distant traffic, no music"      |

If a prompt starts to sprawl, trim the style adjectives first. The subject, action and camera do most of the work.

## Prompting long 20–30 second takes

Seedance 2.5 can render takes of 20 or 30 seconds, which is long enough for a small scene rather than a single moment. The trick is to write those takes as a sequence of beats, in the order they should happen.

A simple way to do it:

1. **Set the opening.** Describe where we start: the subject, the setting and the first framing.
2. **List the beats in order.** Use words like "first", "then", "after that", "finally". Each beat should be one clear change: a movement, a reveal, a line of dialogue.
3. **Give the camera a path.** Say how the camera moves across the beats, for example a slow push-in during the first half and an orbit during the second.
4. **Describe the ending.** Tell the model where the shot should land, so the final seconds do not drift.

As a rough guide, three to five beats fit comfortably in a 20 or 30 second take. Packing ten actions into one clip tends to make each of them rushed. If you need more, render two clips and cut them together.

Before you commit to a long 1080p or 4K render, try the same beats at 480p, or in Seedance 2.5 Flash, to check the pacing. You can compare models in our [Seedance 2.5 vs 2.5 Flash vs 2.0](https://seedance25.com/blog/seedance-2-5-vs-2-5-flash-vs-2-0) breakdown.

## Using reference images

Reference mode lets you upload up to **30 reference images** with Seedance 2.5. That is enough to pin down a character from several angles, a product, a wardrobe, a location and a colour palette all at once.

To get consistent results:

- **Give each image a role.** In the prompt, say what each reference is for: "the woman from the first reference image", "the jacket from the second image", "the street from the third image".
- **Group related images.** If you upload four angles of the same character, say they show one person.
- **Say when an image is only for mood.** "Use the last image only for colour and lighting" stops the model from copying its content.
- **Do not re-describe everything.** The model can see the images. Spend your words on what happens, how the camera moves and what we hear.

More images are not always better. Use as many as the scene really needs, and make sure they agree with each other on things like outfit and lighting.

## Using first and last frame

In first and last frame mode you provide an opening image and, optionally, a closing image. Seedance 2.5 then animates the journey between them.

This is ideal for transformations, reveals and product shots where you know exactly how the clip must start and end. A few tips:

- **Describe the motion, not the frames.** The prompt should explain what happens between the two images: "the camera pulls back as the lights in the building switch on one by one".
- **Keep the two frames compatible.** Same subject, similar lighting and a plausible path between them gives smooth results. Very different images can produce morph-like transitions.
- **Match duration to the change.** A small change fits a short clip. A big change, like day turning to night, needs more seconds.

## When to turn audio off

Seedance 2.5 generates native audio by default, and you can switch it off in the generator. The price stays the same either way, so the choice is purely creative.

Switch audio off when:

- you plan to add your own music or voiceover in an editor,
- the clip will be used as B-roll under a separate soundtrack,
- you are making silent loops for a website or presentation.

Leave it on when the sound is part of the idea, such as dialogue, footsteps, crowd noise or a specific ambience. When audio is on, describe it in the prompt. If you want only natural sound, write "no music".

## 6 copy-paste Seedance 2.5 prompts

Swap in your own subjects and details. Durations and ratios are suggestions.

### 1. Morning café (10 s, 16:9)

```text
A barista with rolled-up sleeves pours steamed milk into a cup, drawing a leaf pattern in the foam. Close-up from the side, then a slow pull-back to reveal a small, sunlit café with plants in the window. Soft morning light, warm tones. Sound of the espresso machine, cups clinking, quiet chatter, no music.
```

### 2. Mountain sunrise long take (30 s, 21:9)

```text
First, a wide static shot of a mountain valley in blue pre-dawn light, mist resting on a lake. Then the first sunlight touches the highest peak and slowly spreads down the slopes. After that, the camera begins a slow push-in over the water as birds take off from the shore. Finally, the sun clears the ridge and the lake turns gold. Cinematic, natural colours. Gentle wind, birdsong, a soft orchestral swell near the end.
```

### 3. Product turntable (8 s, 1:1)

```text
A matte black wireless speaker sits on a pale concrete plinth. The camera orbits slowly around it while a narrow strip of light glides across its surface, revealing the fabric texture. Clean studio backdrop, soft shadows, premium commercial look. Audio off.
```

### 4. Vertical street dance (15 s, 9:16)

```text
Vertical shot of a dancer in a bright orange jacket on a wet city street at night. She starts still, then breaks into a fast footwork routine as neon signs reflect in the puddles. The camera circles her at waist height, then rises to a high angle for the final pose. Upbeat electronic track with a strong beat, splashing footsteps.
```

### 5. Two-line dialogue (10 s, 16:9)

```text
An old man and his granddaughter sit on a wooden porch at dusk. She points at the sky and says, "Is that the first star?" He smiles and answers, "Make a wish before it disappears." Static medium shot, warm lamp light from the doorway, crickets in the background, no music.
```

### 6. Reference character scene (20 s, 16:9)

```text
The young man from the first three reference images, wearing the coat from the fourth image, walks through the train station from the fifth image. First he checks the departure board, then he spots someone off-screen and starts to run. The camera tracks beside him through the crowd and stops as he reaches the platform. Use the last image only for colour. Station announcements, footsteps, rising strings.
```

## Common mistakes to avoid

- **Too many actions in a short clip.** A 4 or 5 second clip has room for one main action.
- **No camera direction.** Always name a move, even if it is "static shot".
- **Beats out of order.** In long takes, write events in the order you want to see them.
- **References without roles.** Every uploaded image should have a clear job in the prompt.
- **Contradictory style.** Pick one look and stay with it.

## Next steps

Open the [generator](https://seedance25.com/#generator), paste one of the prompts above and check the credit cost on the Generate button before you render. If you are new to the studio, start with our step-by-step [how to use Seedance 2.5](https://seedance25.com/blog/how-to-use-seedance-2-5) guide, and see [pricing](https://seedance25.com/pricing) for credit packs and plans.

---

Source: https://seedance25.com/blog/seedance-2-5-prompt-guide · Site index for AI agents: https://seedance25.com/llms.txt · Full content: https://seedance25.com/llms-full.txt
