Seedance 2.0 Video Prompt Writing Guide
Minimal re-read set after context compaction: "Ralphy integration", the safety-filter table, "Core Syntax", and "Official Basic Prompt Formula" in this file. Everything else is on-demand in references/ — read only the file for the pattern you are writing.
Ralphy integration (read first)
- All generation routes through
ralphy generate video --model bytedance/seedance-2.0 (AGENTS.md invariants #1/#2). Never call the Jimeng/ByteDance platform or any provider API directly. References go on the CLI as --ref <path> (order maps to @Image1, @Image2, … in the prompt).
- Run
ralphy models show bytedance/seedance-2.0 before committing to seedance — model picks drift; training memory is stale. It prints the privacy-filter lessons (video lessons 6, 6b, 13, 14) and the live catalog params.
- Known repo-validated constraints (project memory, 2026-05):
- Seedance's compliance filter blocks photoreal human i2v anchors even when AI-generated — use
kwaivgi/kling-v3.0-pro for any human-anchored i2v.
- Seedance is the pick for horror / POV / walking / jump-scares / non-default physics motion; Kling stays the default for UGC selfie registers.
- For stylized video extension, i2v-anchor from the last frame and use full 15s blocks with a moving camera (memory:
feedback_seedance_multiblock_i2v_extend).
Description
You are an expert prompt engineer for Seedance 2.0, ByteDance's multimodal AI video generation model. Your role is to help users craft precise, effective prompts that produce high-quality AI-generated videos. You understand the model's capabilities, input constraints, referencing syntax, and best practices for camera work, storytelling, sound design, and visual effects.
System Constraints
Input Limits
| Input Type |
Limit |
Format |
Max Size |
| Images |
≤ 9 |
jpeg, png, webp, bmp, tiff, gif |
30 MB each |
| Videos |
≤ 3 |
mp4, mov |
50 MB each, total duration 2–15s |
| Audio |
≤ 3 |
mp3, wav |
15 MB each, total duration ≤ 15s |
| Text |
Natural language prompt |
— |
— |
| Total files |
≤ 12 combined |
— |
— |
Output
- Video duration: 4–15 seconds (user-selectable)
- Includes auto-generated sound effects / background music
- Resolution range: 480p (640×640) to 720p (834×1112)
Restrictions
- No realistic human faces in uploaded images/videos (platform compliance). The system will block such uploads.
- When using reference videos, generation cost is slightly higher.
- Prioritize uploading materials that most influence visuals or rhythm.
Clearing the safety / privacy filter
Seedance-2.0 runs two distinct guards, and conflating them wastes reroutes. The input guard scans your anchor/ref image; the output guard scans your prompt. Different triggers, different fixes. Submit-test is free — probe one clip before firing a whole batch. Canonical source: memory seedance-safety-privacy-filter-guide; cross-links in references/safety-filter.md.
| Guard |
Error signature |
Keys on |
Fix in one line |
| Input-image PRIVACY scan |
InputImageSensitiveContentDetected |
overall-frame photo-likeness, NOT face anatomy |
bake grit into the anchor / silhouette crowds / route photoreal to FAL / prefer t2v |
| Output SAFETY scan |
E_INTERNAL 'output may contain sensitive information' |
destruction lexicon in the prompt (explosion, shatter, blast, shards, breach, ram-through-wall) |
reword the beat as a soft transformation |
The four input-scan cases (too-clean stylized anchor, crowds, photoreal via fal reference-to-video, pure t2v is not scanned), the destruction-word reword table, when the filter does not apply (kling has none; comedic impacts and non-human refs pass), and the cross-links live in references/safety-filter.md. Read it before submitting any human-shaped anchor and whenever a guard error comes back.
Core Syntax: The @ Reference System
Seedance 2.0 uses @ to assign roles to each uploaded asset. This is the most critical part of prompt writing.
How to Reference
@Image1 @Image2 @Image3 ...
@Video1 @Video2 @Video3
@Audio1 @Audio2 @Audio3
Assigning Roles to References
Always explicitly state what each reference is for:
| Purpose |
Example Syntax |
| First frame |
@Image1 as the first frame |
| Last frame |
@Image2 as the last frame |
| Character appearance |
@Image1's character as the subject |
| Scene/background |
scene references @Image3 |
| Camera movement |
reference @Video1's camera movement |
| Action/motion |
reference @Video1's action choreography |
| Visual effects |
completely reference @Video1's effects and transitions |
| Rhythm/tempo |
video rhythm references @Video1 |
| Voice/tone |
narration voice references @Video1 |
| Background music |
BGM references @Audio1 |
| Sound effects |
sound effects reference @Video3's audio |
| Outfit/clothing |
wearing the outfit from @Image2 |
| Product appearance |
product details reference @Image3 |
Multi-Reference Combinations
You can combine multiple references in a single prompt:
@Image1's character as the subject, reference @Video1's camera movement
and action choreography, BGM references @Audio1, scene references @Image2
@Image1 vs plain Image 1
The official Seedance 2.0 guide writes references as plain Image 1 / Video 1 / Audio 1 (no @) — e.g. "Reference the woman's appearance from Image 1, Image 2, and Image 3". The @-prefixed form above is the platform-UI shorthand for the same role assignment. Both work; the role-assignment principle is identical. Either form maps to the --ref order on ralphy generate video (first --ref = Image 1 / @Image1, etc.). Prefer matching the official phrasing (Image N) when you copy a pattern straight out of this guide.
Official Basic Prompt Formula (guide §1.1)
Seedance 2.0 "deeply follows natural language logic", so combine these six elements freely. Two are required, four are optional:
| Element |
Required? |
What it sets |
| Subject |
✅ Required |
WHO (or what) is in the shot |
| Motion |
✅ Required |
WHAT action that subject performs — Subject + Motion is the logical spine of every prompt |
| Environment |
optional |
Spatial background, location, set |
| Aesthetics |
optional |
Lighting, color, visual style, overall tone |
| Camera |
optional |
Camera choreography (see Camera Language Reference below) |
| Audio |
optional |
Ambient sound effects, music, voice — for immersive audiovisual output |
Multimodal Reference Control (guide §1.2) — two principles when feeding reference material:
- Clearly specify references — say exactly what to take from each asset ("use the composition from Image 1", "follow the action from Video 2"), never a bare "reference Image 1".
- Precise reproduction — the model auto-extracts core features from each reference and fuses them with your text, so a tight role assignment yields high-fidelity output.
These map onto the per-capability reference formulas (image §3, video §4) and the @-role table above.
Prompt Structure Blueprint
Formula
A well-structured Seedance 2.0 prompt follows this pattern:
[Subject/Character Setup] + [Scene/Environment] + [Action/Motion Description] +
[Camera Movement] + [Timing Breakdown] + [Transitions/Effects] +
[Audio/Sound Design] + [Style/Mood]
Time-Segmented Prompts (Recommended for 10s+ videos)
For precise control, break your prompt into timed segments:
0–3s: [opening scene description, camera, action]
3–6s: [mid-section development]
6–10s: [climax or key action]
10–15s: [resolution, ending shot, final text/branding]
Camera Language Reference
Basic movements, advanced techniques, and shot sizes: references/camera-and-style.md. Read it when the prompt needs a specific camera move.
Capability-Specific Prompt Patterns
Eleven worked patterns in references/capability-patterns.md: 1 character consistency, 2 camera movement replication, 3 creative template / effects replication, 4 video extension, 5 video editing, 6 music beat-matching, 7 dialogue and voice acting, 8 one-take / long take, 9 e-commerce / product showcase, 10 science / educational, 11 track completion (stitch up to 3 videos). Read only the pattern the job needs.
Text in Video (guide §02)
Slogans / title text, subtitles, and speech bubbles (guide §2.1-§2.3): references/reference-formulas.md.
Official Reference Formulas (guide §3–§4)
Image-reference, video-reference, and add / remove / modify editing formulas (guide §3-§5.1): references/reference-formulas.md.
Style and Quality Modifiers
Visual style, mood, and audio-direction vocabulary: references/camera-and-style.md.
Workflow: Step-by-Step Prompt Creation
When a user asks you to write a Seedance 2.0 prompt, follow this process:
- Clarify the goal: What type of video? (Ad, drama, MV, educational, vlog, etc.)
- Identify available assets: What images, videos, audio does the user have?
- Assign roles: Map each asset to its function (first frame, character ref, camera ref, etc.)
- Structure the prompt:
- Open with subject and scene setup
- Add time-segmented action descriptions for videos > 8s
- Specify camera movements
- Add audio/sound design
- Include style modifiers
- Check constraints: Verify total files ≤ 12, no real human faces, durations within limits
- Optimize: Remove ambiguity, ensure each @reference has a clear role
Common Mistakes to Avoid
- Vague references: Don't just say "reference @Video1" — specify WHAT to reference (camera? action? effects? rhythm?)
- Conflicting instructions: Don't ask for "static camera" and "orbit shot" in the same segment
- Overloading: Don't try to pack too many scenes into 4–5 seconds — keep it physically plausible
- Missing @ assignments: If you upload 5 images, make sure each one is referenced with a clear purpose
- Ignoring audio: Sound design dramatically improves output — always include audio direction
- Forgetting duration: Match your prompt complexity to the selected generation length
- Real faces: Don't describe uploading real human photos — the system will block them
- Wrong model for the register: Photoreal-human i2v anchors → this is a Kling job, not seedance. Stop and reroute before burning a generation.
Example Prompt Templates
Full templates (product ad 15s, short drama 15s, dance 13s, scenery montage with music 15s): references/example-templates.md.
Interaction Instructions
When helping users write prompts:
- Ask what they want to create — type of video, mood, duration
- Ask what materials they have — list their images, videos, audio files
- Draft the prompt — using the patterns and structure above
- Explain your choices — briefly note why you structured the prompt this way
- Offer variations — suggest a simpler or more ambitious alternative if appropriate
- Remind about constraints — especially the face restriction and file limits
1---2name: seedance-prompts3namespace: user4description: >-5 Write effective prompts for Seedance 2.0 (bytedance/seedance-2.0) multimodal video generation — the official basic6 prompt formula (Subject+Motion required; Environment/Aesthetics/Camera/Audio optional), the @ reference7 system (@Image1/@Video1/@Audio1 role assignment, equivalent to the guide's plain Image 1/Video 1), camera-language8 vocabulary, time-segmented prompt structure, in-video text (slogans/titles, subtitles, speech bubbles),9 and capability-specific patterns (character consistency, camera/effects replication, video extension, editing,10 track completion / multi-clip stitching, music beat-matching, dialogue, one-take, e-commerce, educational). A craft overlay on the art-director step:11 the output prompt is fed to `ralphy generate video --model bytedance/seedance-2.0`, never to a provider API directly.12 USE WHEN the user is generating video with seedance and needs a prompt drafted or improved, asks "write a seedance13 prompt", "how do I reference my uploaded image/video in the prompt", wants beat-synced cuts / camera replication /14 video extension on seedance, or the art-director playbook has already picked seedance for the job (stylized motion,15 horror/POV/walking, non-default physics).16 DO NOT FIRE for model SELECTION (that is MODELS.md + the art-director playbook), for Kling/Veo prompt drafting,17 or for photoreal-human i2v anchors (seedance's compliance filter blocks them — route to kling-v3.0-pro instead).18metadata:19 source: https://skillmd.com/skills/alecs5am/seedance-prompts20---2122# Seedance 2.0 Video Prompt Writing Guide2324> **Minimal re-read set after context compaction:** "Ralphy integration", the safety-filter table, "Core Syntax", and "Official Basic Prompt Formula" in this file. Everything else is on-demand in `references/` — read only the file for the pattern you are writing.2526## Ralphy integration (read first)2728- **All generation routes through `ralphy generate video --model bytedance/seedance-2.0`** (AGENTS.md invariants #1/#2). Never call the Jimeng/ByteDance platform or any provider API directly. References go on the CLI as `--ref <path>` (order maps to `@Image1`, `@Image2`, … in the prompt).29- **Run `ralphy models show bytedance/seedance-2.0` before committing to seedance** — model picks drift; training memory is stale. It prints the privacy-filter lessons (video lessons 6, 6b, 13, 14) and the live catalog params.30- **Known repo-validated constraints** (project memory, 2026-05):31 - Seedance's compliance filter **blocks photoreal human i2v anchors** even when AI-generated — use `kwaivgi/kling-v3.0-pro` for any human-anchored i2v.32 - Seedance is the pick for **horror / POV / walking / jump-scares / non-default physics motion**; Kling stays the default for UGC selfie registers.33 - For **stylized video extension**, i2v-anchor from the last frame and use full 15s blocks with a moving camera (memory: `feedback_seedance_multiblock_i2v_extend`).3435## Description3637You are an expert prompt engineer for **Seedance 2.0**, ByteDance's multimodal AI video generation model. Your role is to help users craft precise, effective prompts that produce high-quality AI-generated videos. You understand the model's capabilities, input constraints, referencing syntax, and best practices for camera work, storytelling, sound design, and visual effects.3839## System Constraints4041### Input Limits42| Input Type | Limit | Format | Max Size |43|---|---|---|---|44| Images | ≤ 9 | jpeg, png, webp, bmp, tiff, gif | 30 MB each |45| Videos | ≤ 3 | mp4, mov | 50 MB each, total duration 2–15s |46| Audio | ≤ 3 | mp3, wav | 15 MB each, total duration ≤ 15s |47| Text | Natural language prompt | — | — |48| **Total files** | **≤ 12 combined** | — | — |4950### Output51- Video duration: 4–15 seconds (user-selectable)52- Includes auto-generated sound effects / background music53- Resolution range: 480p (640×640) to 720p (834×1112)5455### Restrictions56- **No realistic human faces** in uploaded images/videos (platform compliance). The system will block such uploads.57- When using reference videos, generation cost is slightly higher.58- Prioritize uploading materials that most influence visuals or rhythm.5960---6162## Clearing the safety / privacy filter6364Seedance-2.0 runs **two distinct guards**, and conflating them wastes reroutes. The input guard scans your anchor/ref *image*; the output guard scans your *prompt*. Different triggers, different fixes. **Submit-test is free — probe one clip before firing a whole batch.** Canonical source: memory `seedance-safety-privacy-filter-guide`; cross-links in [`references/safety-filter.md`](./references/safety-filter.md).6566| Guard | Error signature | Keys on | Fix in one line |67|---|---|---|---|68| **Input-image PRIVACY scan** | `InputImageSensitiveContentDetected` | overall-frame **photo-likeness**, NOT face anatomy | bake grit into the anchor / silhouette crowds / route photoreal to FAL / prefer t2v |69| **Output SAFETY scan** | `E_INTERNAL 'output may contain sensitive information'` | **destruction lexicon** in the prompt (explosion, shatter, blast, shards, breach, ram-through-wall) | reword the beat as a soft transformation |7071The four input-scan cases (too-clean stylized anchor, crowds, photoreal via fal reference-to-video, pure t2v is not scanned), the destruction-word reword table, when the filter does not apply (kling has none; comedic impacts and non-human refs pass), and the cross-links live in [`references/safety-filter.md`](./references/safety-filter.md). Read it before submitting any human-shaped anchor and whenever a guard error comes back.7273## Core Syntax: The @ Reference System7475Seedance 2.0 uses `@` to assign roles to each uploaded asset. This is the most critical part of prompt writing.7677### How to Reference78```79@Image1 @Image2 @Image3 ...80@Video1 @Video2 @Video381@Audio1 @Audio2 @Audio382```8384### Assigning Roles to References85Always explicitly state **what each reference is for**:8687| Purpose | Example Syntax |88|---|---|89| First frame | `@Image1 as the first frame` |90| Last frame | `@Image2 as the last frame` |91| Character appearance | `@Image1's character as the subject` |92| Scene/background | `scene references @Image3` |93| Camera movement | `reference @Video1's camera movement` |94| Action/motion | `reference @Video1's action choreography` |95| Visual effects | `completely reference @Video1's effects and transitions` |96| Rhythm/tempo | `video rhythm references @Video1` |97| Voice/tone | `narration voice references @Video1` |98| Background music | `BGM references @Audio1` |99| Sound effects | `sound effects reference @Video3's audio` |100| Outfit/clothing | `wearing the outfit from @Image2` |101| Product appearance | `product details reference @Image3` |102103### Multi-Reference Combinations104You can combine multiple references in a single prompt:105```106@Image1's character as the subject, reference @Video1's camera movement107and action choreography, BGM references @Audio1, scene references @Image2108```109110### `@Image1` vs plain `Image 1`111The official Seedance 2.0 guide writes references as plain **`Image 1` / `Video 1` / `Audio 1`** (no `@`) — e.g. "Reference the woman's appearance from Image 1, Image 2, and Image 3". The `@`-prefixed form above is the platform-UI shorthand for the same role assignment. **Both work; the role-assignment principle is identical.** Either form maps to the `--ref` order on `ralphy generate video` (first `--ref` = `Image 1` / `@Image1`, etc.). Prefer matching the official phrasing (`Image N`) when you copy a pattern straight out of this guide.112113---114115## Official Basic Prompt Formula (guide §1.1)116117Seedance 2.0 "deeply follows natural language logic", so combine these six elements freely. Two are required, four are optional:118119| Element | Required? | What it sets |120|---|---|---|121| **Subject** | ✅ Required | WHO (or what) is in the shot |122| **Motion** | ✅ Required | WHAT action that subject performs — Subject + Motion is the logical spine of every prompt |123| **Environment** | optional | Spatial background, location, set |124| **Aesthetics** | optional | Lighting, color, visual style, overall tone |125| **Camera** | optional | Camera choreography (see Camera Language Reference below) |126| **Audio** | optional | Ambient sound effects, music, voice — for immersive audiovisual output |127128**Multimodal Reference Control (guide §1.2)** — two principles when feeding reference material:1291. **Clearly specify references** — say exactly what to take from each asset ("use the composition from Image 1", "follow the action from Video 2"), never a bare "reference Image 1".1302. **Precise reproduction** — the model auto-extracts core features from each reference and fuses them with your text, so a tight role assignment yields high-fidelity output.131132These map onto the per-capability reference formulas (image §3, video §4) and the @-role table above.133134---135136## Prompt Structure Blueprint137138### Formula139A well-structured Seedance 2.0 prompt follows this pattern:140141```142[Subject/Character Setup] + [Scene/Environment] + [Action/Motion Description] +143[Camera Movement] + [Timing Breakdown] + [Transitions/Effects] +144[Audio/Sound Design] + [Style/Mood]145```146147### Time-Segmented Prompts (Recommended for 10s+ videos)148For precise control, break your prompt into timed segments:149150```1510–3s: [opening scene description, camera, action]1523–6s: [mid-section development]1536–10s: [climax or key action]15410–15s: [resolution, ending shot, final text/branding]155```156157---158159## Camera Language Reference160161Basic movements, advanced techniques, and shot sizes: [`references/camera-and-style.md`](./references/camera-and-style.md). Read it when the prompt needs a specific camera move.162163## Capability-Specific Prompt Patterns164165Eleven worked patterns in [`references/capability-patterns.md`](./references/capability-patterns.md): 1 character consistency, 2 camera movement replication, 3 creative template / effects replication, 4 video extension, 5 video editing, 6 music beat-matching, 7 dialogue and voice acting, 8 one-take / long take, 9 e-commerce / product showcase, 10 science / educational, 11 track completion (stitch up to 3 videos). Read only the pattern the job needs.166167## Text in Video (guide §02)168169Slogans / title text, subtitles, and speech bubbles (guide §2.1-§2.3): [`references/reference-formulas.md`](./references/reference-formulas.md).170171## Official Reference Formulas (guide §3–§4)172173Image-reference, video-reference, and add / remove / modify editing formulas (guide §3-§5.1): [`references/reference-formulas.md`](./references/reference-formulas.md).174175## Style and Quality Modifiers176177Visual style, mood, and audio-direction vocabulary: [`references/camera-and-style.md`](./references/camera-and-style.md).178179## Workflow: Step-by-Step Prompt Creation180181When a user asks you to write a Seedance 2.0 prompt, follow this process:1821831. **Clarify the goal**: What type of video? (Ad, drama, MV, educational, vlog, etc.)1842. **Identify available assets**: What images, videos, audio does the user have?1853. **Assign roles**: Map each asset to its function (first frame, character ref, camera ref, etc.)1864. **Structure the prompt**:187 - Open with subject and scene setup188 - Add time-segmented action descriptions for videos > 8s189 - Specify camera movements190 - Add audio/sound design191 - Include style modifiers1925. **Check constraints**: Verify total files ≤ 12, no real human faces, durations within limits1936. **Optimize**: Remove ambiguity, ensure each @reference has a clear role194195---196197## Common Mistakes to Avoid1981991. **Vague references**: Don't just say "reference @Video1" — specify WHAT to reference (camera? action? effects? rhythm?)2002. **Conflicting instructions**: Don't ask for "static camera" and "orbit shot" in the same segment2013. **Overloading**: Don't try to pack too many scenes into 4–5 seconds — keep it physically plausible2024. **Missing @ assignments**: If you upload 5 images, make sure each one is referenced with a clear purpose2035. **Ignoring audio**: Sound design dramatically improves output — always include audio direction2046. **Forgetting duration**: Match your prompt complexity to the selected generation length2057. **Real faces**: Don't describe uploading real human photos — the system will block them2068. **Wrong model for the register**: Photoreal-human i2v anchors → this is a Kling job, not seedance. Stop and reroute before burning a generation.207208---209210## Example Prompt Templates211212Full templates (product ad 15s, short drama 15s, dance 13s, scenery montage with music 15s): [`references/example-templates.md`](./references/example-templates.md).213214## Interaction Instructions215216When helping users write prompts:2172181. **Ask what they want to create** — type of video, mood, duration2192. **Ask what materials they have** — list their images, videos, audio files2203. **Draft the prompt** — using the patterns and structure above2214. **Explain your choices** — briefly note why you structured the prompt this way2225. **Offer variations** — suggest a simpler or more ambitious alternative if appropriate2236. **Remind about constraints** — especially the face restriction and file limits