skills/krea-generate/references/models/seedance-2.md

---
name: seedance-2-prompting
description: Prompting playbook for the Seedance video family (Seedance 2.5, 2.0, Fast, Mini), focused on multimodal references, shot sequencing, action/camera control, audio/text, video editing, extension, and Krea operating rules.
---

# Seedance Prompting Guide

Load this file only after the selected model is Seedance 2.5 (`bytedance/seedance-2-5`), Seedance 2.0, Seedance 2 Fast/Mini, or a closely related Seedance video model. Seedance 2.5 is the flagship; the whole family shares the prompt language below.

## Prompting Stance

You are an expert Seedance video prompt engineer. Convert the user's intent into model-native shot instructions: explicit asset roles, stable subject definitions, concrete motion, one camera move per shot, audio/text rules, and a constraints tail. Prefer physically plausible staged action over vague cinematic adjectives.

Use the rules in this file as the controlling guidance for Seedance prompting. Example patterns in the appendix are optional support material and should not override the main rules.

Always verify the live Krea model schema before submitting. Use only fields exposed by the selected model. Prompt examples describe intent; schema controls such as duration, aspect, resolution, references, audio, `start_image`, `end_image`, and reference arrays must come from the live schema.

## Examples Appendix

For concrete prompt examples, search `seedance-2-examples.md` by heading and read only the matching section:

```bash
rg -n "^##|^###" seedance-2-examples.md
rg -n "Multimodal Reference|Video Editing|Shot Storyboard" seedance-2-examples.md
```

Do not load the full examples appendix unless the relevant section cannot be identified.

## When To Load Examples

| Task | Search terms |
|---|---|
| image/video/audio references | `Multimodal Reference|Asset Role` |
| subject identity from references | `Subject Definition|Multi-Subject Reference` |
| multi-shot story or drama | `Shot Storyboard|Dialogue Short Drama` |
| action/fight/chase | `Action Scene|Wuxia Confrontation|Transformation Arc` |
| video edit, add/remove/replace | `Video Editing|Object Removal|Replace Product` |
| video extension or track connection | `Video Extension|Track Completion` |
| text, subtitles, speech bubbles | `Text Generation|Subtitles|Speech Bubble` |
| video/camera/effects reference | `Video Motion Reference|Special Effects Reference` |
| voice, dialogue, audio reference | `Voice Timbre|Dialogue Audio` |
| product ad or commercial | `Product Commercial` |
| stylized animation | `Stylized Animation` |
| failure/retake | `Retake Examples|Troubleshooting` |

## Core Formula

Seedance is a multimodal video model that uses text, images, video, and audio together. Write prompts as engineering-style shot instructions, not copywriting.

Advanced formula:

```text
precise subject + action details + scene/environment + lighting/color tone
+ camera movement + visual style + constraints
```

For complex videos, use a storyboard:

```text
References: <asset roles>
Global style: <visual tone, constraints>
Shot 1: <camera/transition> + <subject action/expression> + <position/space> + <audio>
Shot 2: <camera/transition> + <subject action/expression> + <position/space> + <audio>
Constraints: <identity, motion stability, no subtitles/logos/watermarks, style locks>
```

## Reference Roles

Name what each asset contributes. Do not assume the model knows why an asset was uploaded.

| Asset role | Prompt pattern |
|---|---|
| Image subject | `Use the <2-3 stable traits> in @Image1 as <Subject_A>.` |
| Image scene/style | `Use @Image2 as the scene/style reference, not as a subject.` |
| Video action | `Reference the <action/motion> in @Video1.` |
| Video camera | `Reference the <camera movement/shot language> in @Video1.` |
| Video effects | `Reference the <effect form and motion logic> in @Video1.` |
| Audio timbre | `Use the <voice traits> timbre from @Audio1 for <speaker>.` |
| Audio rhythm | `Use @Audio1 for background rhythm / atmosphere.` |

Subject definitions should be concise and stable: 2-3 clear visual traits, no contradictions. If an asset has multiple people/props, label the intended subject explicitly. Even when using Krea asset IDs, prompt with `@ImageN`, `@VideoN`, or `@AudioN` labels because those are what the model understands in the prompt.

Recommended asset strategy: use the fewest assets that carry the job, usually 4-5 total: 1-2 character images, 1 scene image, 1 camera/action video, and 1 audio clip. Too many assets can confuse priority, blur subject identification, and cause style conflicts.

## Task Formulas

### Multimodal Reference

Use when extracting subject, style, scene, action, camera movement, sound effect, or voice from references to make a new video.

```text
Reference <dimension> from @ImageN/@VideoN/@AudioN to generate <new scene>.
Keep <specific traits> consistent. Change <new action/environment/story>.
```

### Video Editing

Use when modifying an existing video. Parts not mentioned should remain unchanged.

```text
Strictly edit @Video1. Change only <target>.
Add: <element features> + <timing> + <location>.
Remove: <element>, keeping all other video content unchanged.
Replace: <old element> with <new element/reference>, preserving motion and camera work.
```

### Video Extension

Use when continuing or prepending a single coherent scene.

```text
Extend @Video1 forward/backward: <what happens before/after>.
Preserve continuity of <characters, lighting, camera, action path, audio tone>.
```

Extension is best for single-scene dialogue, emotional progression, or movement along one path. Use segmented stitching for plot turns, chases, fights, montages, or complex action.

### Track Completion / Stitching

Use when connecting multiple input clips:

```text
@Video1 + <transition description> + @Video2 + <transition description> + @Video3.
```

Respect the live schema. For the 2.0 family, official guidance notes a maximum of 3 input video clips and total input video duration not exceeding 15 seconds for track-completion style inputs. Seedance 2.5 accepts up to 10 video references with a ~15s cap per clip rather than in aggregate.

## Shot Sequencing

Prefer `Shot 1`, `Shot 2`, `Shot 3` over strict second-by-second timing. Precise timing such as `0-3s` can be unstable. Use time ranges only as editorial staging when a workflow specifically requires them; do not expect frame-accurate timing.

Each shot should include:

1. camera movement or transition
2. subject action and expression
3. position or spatial relationship
4. audio, dialogue, ambience, or sound effect

For multi-shot creator-format prompts, put the shot count near the top as metadata. Keep this as metadata, not a substitute for shot details. Aspect ratio, total duration, and resolution are controls, not prompt content — set them through the live schema fields (see the stance note at the top), never in the prompt text.

## Action And Camera Rules

Actions need body-part specificity, degree, speed, force, and continuity:

- `she slowly raises her right hand`
- `he turns his head quickly toward the doorway`
- `the runner pushes hard off the wet pavement`
- `shoulders relax, breath releases, faint smile appears`

Prefer slow, gentle, coherent small movements for stability. Avoid stacking high-burst actions unless you split them into separate shots or clips.

Camera rules:

- Use standard terms: medium shot, close-up, wide shot, fixed shot, slow push-in, smooth lateral track.
- Specify only one primary camera movement per shot.
- Do not ask for push, pull, pan, orbit, and crane in the same shot.
- For action that travels, name direction, distance, start/end pose, and frame position.

## Cinematic Cut Sequences

Krea house doctrine for staged multi-beat work: product reveals, logo stings, macro studies, object films. It overrides three rules above; everything else in this file still applies. For the craft behind it — what to stage, which reveal to pick, how to light it — use the `krea-motion` skill.

Three overrides:

1. **Beats carry explicit time ranges**, not `Shot 1` / `Shot 2` labels. Unstaged boundaries render as dissolves instead of cuts. Write `SHOT 2 — EDGE (1.4-2.6s)`.
2. **Pace is quantified, never described.** `slow`, `gently`, `softly`, `dreamy float` read as *slow down the footage* and come back speed-ramped. Give distance, duration and a constant rate: `one continuous push-in traveling 20cm across 6 seconds, motion-control smooth, never accelerating`. Slowness belongs to the camera; the physics stay realtime.
3. **A fully staged prompt holds four beats**, not three — but only when every beat carries its own time range, shot size, lens, one camera move, one physical event and a named transition out. A fifth beat merges into the fourth. This supersedes the three-beat figure in the Krea live test supplement below for staged work; three remains the right expectation for an unstaged prompt.

Four blocks these depend on, in prompt order:

- **World** — one environment, one dominant light, one palette, declared once and held across every beat. Without it, beat 3 relocates.
- **Motion split** — state separately what is nailed down, what the camera does, and what moves at realtime speed. The highest-leverage block and the most skipped: `the object is locked, no rotation, wobble, drift or scale change unless a beat says so; camera moves are quantified per beat; atmospherics move at realtime speed; no speed ramping`.
- **Light** — name the source, its behaviour, and hard negatives. The untamed default is god rays, flare and a blue push: `NO god rays, NO lens flare, NO bloom, NO blue filter`.
- **Constraints tail** — every lock restated at the end. The last 20% of a long generation drifts back toward defaults; the tail is the anti-drift mechanism, not redundancy.

Beat shape:

```text
SHOT 1 — <NAME> (0-Xs), <size>, <lens>. <one camera move>. <one physical event>.
  <named transition>.
...
SHOT 4 — <NAME> (Z-Ws), full frame, <lens>. <the whole subject, first and only
  time>. <one camera move>. <holds still>.
```

Never write two camera moves in one beat. "Slowly orbit while pushing in and tilting up" is three instructions and returns a lurch.

Retakes change **one** block per run with the seed pinned. Changing three at once teaches you nothing about which one mattered.

## Text, Dialogue, And Audio

Seedance can generate common text such as slogans, subtitles, and speech bubbles. Use simple/common characters and avoid rare symbols unless needed.

Text prompt pattern:

```text
<Text content> + <timing> + <position> + <entrance/appearance> + <visual attributes>
```

Dialogue language should be consistent; avoid mixing languages except for proper nouns.

Special symbol conventions:

| Information type | Symbol style |
|---|---|
| Music | `(fast-paced drums in the background)` |
| Sound effect | `<dog barking in the distance>` |
| Dialogue | `{Hello, world}` |
| Dialogue in less common language | `says in Japanese {こんにちは}` |
| Subtitle text | `〖Chapter One: Departure〗` |

Voice matching improves when you describe voice traits, such as age, thickness, warmth, grain, pace, and emotion, rather than only saying "use @Audio1."

## Krea Operating Rules

These are Krea-specific rules for the Seedance family (`bytedance/seedance-2-5`, `bytedance/seedance-2`, and the Fast/Mini variants); live schema still wins.

Variant capabilities, as of the last live check:

| Variant | Duration | Resolutions | Reference budget |
|---|---|---|---|
| `bytedance/seedance-2-5` | 4-30s | 480p/720p | 50 total: 30 images / 10 videos / 10 audio |
| `bytedance/seedance-2` | 4-15s | 480p/720p/1080p/4k | 12 total: 9 images / 3 videos / 3 audio |
| `bytedance/seedance-2-fast`, `-mini` | 4-15s | 480p/720p | 12 total: 9 images / 3 videos / 3 audio |

Seedance 2.5 is the delivery default. Reach for Seedance 2.0 only when the deliverable needs 1080p or 4k, which 2.5 does not expose. Start/end frames count toward the image-reference budget when mixed with references.

- `@Image1` / `@Video1` / `@Audio1` are prompt labels. Also pass assets through the schema-confirmed fields.
- In Krea, treat `end_image` and `reference_images` as mutually exclusive for Seedance.
- Chained/destination shots use `start_image` + `end_image`, omitting `reference_images`.
- Terminal/detail-anchored shots use `start_image` + `reference_images`, omitting `end_image`.
- `end_image` is a visual destination, not a loose style reference. Keep it within plausible story-time from the start frame.
- Seedance has a 4 second minimum duration. For shorter editorial cuts, generate 4s and trim if useful motion lands early.
- For in-scene continuity, extract the prior clip's actual last frame and use it as the next `start_image`.
- Treat completed jobs with no result URL as failure/refusal, not success. Retry once with a sanitized prompt — drop proper nouns, IP-like phrases and inessential signage text, keep the `start_image`. If it fails again, drop `end_image` and retry start-image-only.
- Submit video jobs in waves of 12 or fewer to avoid practical workspace concurrency limits.
- `enhance_prompt` off for authored work. When you have written the prompt precisely, a rewriter flattens it; leave it on only when the user hands over a one-line brief and wants a fast look.
- Reach for 21:9 on cinematic reveals.
- `generate_audio` is not a real choice on Seedance: every variant, 2.5 included, forces it on, so the submitted value is discarded and every take returns a generated bed. Keep an `Audio:` block anyway to steer what that bed contains, and strip the track at delivery with `ffmpeg -i take.mp4 -c:v copy -an out.mp4` when the piece needs silence or a designed bed.

Krea live test supplement: one-shot commercial timelines usually land about 3 strong beats reliably. More beats need full staging, short beats, and a constraints tail; unstaged "hard cuts" can morph instead of cutting.

## Constraints And Troubleshooting

Add constraints at the tail, especially for longer prompts:

- `keep it subtitle-free`
- `avoid generating any text or subtitles`
- `do not generate a logo`
- `do not generate a watermark`
- `faces and body proportions remain stable`
- `motion is continuous and natural, no stutter, no flicker, no clipping`
- `do not duplicate characters or create twin avatars`

Known fixes:

| Symptom | Fix |
|---|---|
| Unwanted subtitles | Add subtitle-free constraints, remove text from references, prefer landscape if feasible |
| Logo/watermark appears | Add explicit no-logo / no-watermark constraints |
| Style drifts to realism | State the target style; if needed, convert references to the target style before video generation |
| Extension join jumps | Repair in post by aligning keyframes; trim a few frames at joins when needed |
| Duplicate/twin character appears | Use single-person refs, simplify prompt, add anti-duplicate constraint |
| Extension quality degrades | Limit stacked continuations, use high-definition refs, consider a white-model intermediate for long continuation workflows |
| Special effect is wrong | Use a reference video that shows the effect form and motion logic |
| Voice match is weak | Describe voice traits explicitly in addition to attaching @Audio |
| Motion is unstable | Reduce action intensity, use one camera move, break action into shots |
| Everything looks underwater | No motion split, or pace adverbs instead of numbers. Add the motion split, lock the subject, state realtime physics |
| Beat 3 is in a different room | No world block. Declare one world and restate it in the constraints tail |
| Hard cuts became dissolves | Beats are under-staged. Give each one a time range, size, lens, move, event and named transition — not a louder HARD CUT |
| Logo re-lettered or warped | No text lock. "keep every printed line exactly as shown, do not re-letter or warp", in the references block AND the tail |
| Random god rays and flare | Light source unnamed. Add the light block with explicit negatives |
| Camera lurches | Two moves in one beat. One only |
| Subject rotates for no reason | Subject not locked in the motion split |
| Back half goes lazy | Long final hold. Cap beats and add "no long static holds, keep cutting" |
| Great first 2s, then drifts | No constraints tail. Restate every lock at the end |
| Stock trailer music appears | No audio direction. Name the diegetic sound or ban music explicitly |