---
name: seedance-2-5-video-gen
description: "Build, validate, dry-run, and execute all five Seedance 2.5 workflows through EvoLink: text-to-video, image-to-video, reference-to-video, video edit, and video extend. Use when a user wants a generation-ready Seedance 2.5 prompt, payload, API workflow, edit, extension, or reference-media task. Do not trigger for general model news, pricing questions, or unrelated video editing advice."
homepage: https://github.com/cheercheung/seedance-2-5-video-generate-api-skill
metadata: {"openclaw":{"homepage":"https://github.com/cheercheung/seedance-2-5-video-generate-api-skill","requires":{"bins":["jq","curl"],"env":["EVOLINK_API_KEY"]},"primaryEnv":"EVOLINK_API_KEY"}}
---

# Seedance 2.5 Video Generation

Turn rough video intent into a coherent Seedance 2.5 prompt and a validated EvoLink API request for text, image, reference generation, video editing, or video extension

> Availability boundary: check all five current official documentation routes before any live task. A valid schema or successful dry run does not prove account access or production availability

## When to Activate This Skill

Activate when the user asks to:

- Create or improve a Seedance 2.5 video prompt
- Build a Seedance 2.5 API payload
- Generate video from text
- Animate one image or a first/last-frame pair
- Use images, videos, or audio as references
- Edit or extend a reference video
- Plan one continuous shot or a short multi-shot sequence
- Run, dry-run, troubleshoot, or integrate an EvoLink Seedance 2.5 request

Do not activate for:

- Model news, benchmarks, availability, or pricing questions with no generation task
- Generic video editing that does not use Seedance 2.5
- Requests to fabricate missing reference assets
- Requests that require guaranteed identity, logo, or exact frame-level consistency

## Supported Workflows

| Mode | Model ID | Required media | Unsupported input |
|---|---|---|---|
| `text` | `seedance-2.5-text-to-video` | None | Image, video, and audio URL arrays |
| `image` | `seedance-2.5-image-to-video` | One or two `image_urls` | Video and audio URL arrays; web search |
| `reference` | `seedance-2.5-reference-to-video` | At least one image, video, or audio URL | Edit/extend intent, web search, and Base64 media |
| `edit` | `seedance-2.5-video-edit` | One to ten `video_urls`; first video is the edit target | Custom duration, fixed aspect ratio, web search, and Base64 media |
| `extend` | `seedance-2.5-video-extend` | One to ten `video_urls`; first video is the extension target | Fixed aspect ratio, web search, and Base64 media |

Reference prompts should identify materials explicitly as `@image1`, `@video1`, and `@audio1`

## Script Location

Resolve every path relative to the directory containing this `SKILL.md`:

```text
SKILL_DIR = directory containing this SKILL.md
SCRIPT = {SKILL_DIR}/scripts/seedance-2-5-video-gen.sh
```

## After Installation

1. Check whether `EVOLINK_API_KEY` is set without printing its value
2. If missing, open or show `https://evolink.ai/dashboard/keys?utm_source=skill&utm_medium=install&utm_campaign=seedance-2-5-video-gen`
3. Ask the user to create or select a key and paste it back
4. Save it as `EVOLINK_API_KEY` for the current session and offer persistent shell storage
5. Validate it with the non-generating `GET /v1/credits` path before a real task
6. Run the availability probe and explain the current official state

Commands:

```bash
export EVOLINK_API_KEY="user_key"
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh --validate-key
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh --check-availability
```

Never expose the key in logs or repeat it back to the user

## Core Principles

1. Preserve user intent, approved assets, duration, aspect ratio, dialogue, and silence constraints
2. Design motion over time: opening state, main action, camera behavior, background motion, and ending state
3. Keep the action physically plausible for 4 to 30 seconds and avoid cramming many scenes into a short clip
4. Separate visual direction from dialogue, sound effects, ambience, and music
5. Prefer one coherent shot unless the user explicitly requests a sequence and the duration can support it
6. Describe continuity targets without promising perfect consistency or frame-exact control
7. Validate the payload before any call and never retry after `TASK_SUBMITTED` unless the user explicitly approves a new paid task

## Flow

### Step 1: Understand the Deliverable

Extract:

- Intended use: ad, product demo, cinematic shot, social clip, edit, extension, or other
- Duration and aspect ratio
- Subject and environment
- Opening frame, temporal action, and ending frame
- Camera movement and stability
- Audio, dialogue, sound effects, ambience, or silence
- Reference assets and the intended role of each
- Output preference: prompt, Markdown plan, JSON payload, dry run, or live task

### Step 2: Choose the Mode

- No media input: `text`
- One or two images that define the first frame or first/last frames: `image`
- Style, identity, product, motion, or audio references used to generate new footage: `reference`
- Change the content of an existing video while preserving its timeline: `edit`
- Continue an existing video forward or backward: `extend`

Do not send edit or extension intent to `reference`: the API can accept the request initially and then fail it during model execution

If the mode is ambiguous, ask one compact question that includes all missing required information

### Step 3: Build the Video Prompt

Use this order:

1. Shot format and duration
2. Subject, wardrobe/product details, and environment
3. Opening state
4. Main action and temporal progression
5. Camera framing and movement
6. Background and secondary motion
7. Lighting, color, texture, and mood
8. Continuity constraints
9. Ending state
10. Audio instructions, kept distinct from the visual description

For short clips, use one primary subject action and one camera move. Dialogue must fit naturally within the duration

### Step 4: Validate Model Constraints

Shared defaults and ranges:

- `duration`: 5; integer range 4 to 30, or `-1` for automatic duration
- `quality`: `720p`, also supports `480p`
- `aspect_ratio`: `adaptive`, `16:9`, `9:16`, `1:1`, `4:3`, `3:4`, or `21:9`, subject to mode rules below
- `output_format`: `mp4` or `mov`
- `generate_audio`: true
- `content_filter`: true

Mode rules:

- Text mode may use `model_params.web_search`
- Image mode requires one or two public JPEG, PNG, or WebP URLs and only accepts `aspect_ratio=adaptive`
- Reference mode accepts up to 30 images, 10 videos, and 10 audio clips, with at least one total input, and is only for generating new footage
- Edit mode requires one to ten video URLs, uses the first as the edit target, accepts only `duration=-1` and `aspect_ratio=adaptive`, and allows supplementary image/audio references
- Extend mode requires one to ten video URLs, uses the first as the extension target, accepts `duration=-1` or 4 to 30 and only `aspect_ratio=adaptive`, and allows supplementary image/audio references
- Media must be directly accessible URLs, not local paths or Base64 data
- Callback URLs must use HTTPS and must not point to private networks

### Step 5: Preview Before Spending Credits

Default to a dry run until the user approves the exact payload and cost boundary:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "A red T-shirt rotates once on a seamless white studio set; slow dolly-in; fabric remains stable; no dialogue, soft room tone" \
  --mode text --duration 5 --aspect-ratio 1:1 --dry-run
```

Show the user the chosen model, prompt, media list, duration, quality, aspect ratio, audio setting, and callback behavior

### Step 6: Execute Once When Available and Approved

Before a real task:

- Confirm official availability
- Confirm the key passed non-generating validation
- Confirm the user-approved cost boundary
- Confirm this exact payload

Run the script once. Poll until `completed` or `failed`, or return after task creation when a callback is supplied

## Script Usage

Text-to-video dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "Single continuous shot: a paper airplane crosses a sunlit studio, the camera pans smoothly to follow it, dust motes drift, it lands centered on a wooden desk" \
  --mode text --duration 5 --aspect-ratio 16:9 --dry-run
```

Image-to-video dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "The product remains unchanged while the camera makes a subtle premium push-in; soft highlights travel across the bottle; quiet ambience, no dialogue" \
  --mode image --image-url https://images.unsplash.com/photo-1523275335684-37898b6baf30?w=1200 --duration 5 --dry-run
```

First/last-frame dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "Transition naturally from the first frame to the last frame; preserve the character, clothes, and lighting; slow clockwise camera arc" \
  --mode image \
  --image-url https://images.unsplash.com/photo-1523275335684-37898b6baf30?w=1200 \
  --image-url https://images.unsplash.com/photo-1542291026-7eec264c27ff?w=1200 \
  --duration 6 --dry-run
```

Reference-to-video dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "Generate a new shot using @video1 for camera motion, @image1 for product identity, and @audio1 as background music" \
  --mode reference \
  --image-url https://images.unsplash.com/photo-1523275335684-37898b6baf30?w=1200 \
  --video-url https://storage.googleapis.com/gtv-videos-bucket/sample/ForBiggerBlazes.mp4 \
  --audio-url https://samplelib.com/lib/preview/mp3/sample-3s.mp3 \
  --duration 8 --dry-run
```

Video edit dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "Edit @video1: replace the cup with the product shown in @image1 while preserving timing, camera motion, and background" \
  --mode edit \
  --video-url https://storage.googleapis.com/gtv-videos-bucket/sample/ForBiggerBlazes.mp4 \
  --image-url https://images.unsplash.com/photo-1523275335684-37898b6baf30?w=1200 \
  --duration -1 --dry-run
```

Video extend dry run:

```bash
{SKILL_DIR}/scripts/seedance-2-5-video-gen.sh \
  "Extend @video1 forward: continue the same camera movement and subject motion, preserve lighting and identity, end on the product logo" \
  --mode extend \
  --video-url https://storage.googleapis.com/gtv-videos-bucket/sample/ForBiggerBlazes.mp4 \
  --duration 8 --dry-run
```

## Output Format

When the user asks for a prompt or plan instead of execution, return:

```md
## Video Prompt

[Copy-ready prompt]

## Scene Structure

- Mode:
- Duration:
- Aspect ratio:
- Shot type:
- Opening state:
- Main action:
- Camera movement:
- Subject movement:
- Background movement:
- Ending state:

## Audio Instructions

[Dialogue, sound effects, ambience, music, or silence]

## Continuity Constraints

[Details that should remain stable]

## API Settings

[Model ID and supported request fields]
```

When JSON is requested, return valid JSON that uses only official request fields

## Script Output Protocol

| Output | Meaning | Agent action |
|---|---|---|
| `DRY_RUN=true` | Payload was validated without an API call | Show the payload and state that zero credits were spent |
| `KEY_VALID=true` | Non-generating key validation passed | Continue to availability and cost confirmation |
| `launch_status=closed` | Official docs still mark a route unavailable | Do not make a generation call |
| `TASK_SUBMITTED: task_id=...` | A paid task is already queued | Do not resubmit automatically |
| `STATUS_UPDATE: ...` | Poll progress | Relay useful progress without spamming |
| `RESULT_URL=...` | Final asset URL | Return it and recommend saving it promptly |
| `CALLBACK_PENDING: ...` | Callback mode accepted | Return the task ID and callback URL |
| `POLL_TIMEOUT: ...` | Local polling ended, task may continue | Return the task ID; do not retry blindly |
| `ERROR: ...` | Validation or API failure | Explain the exact correction needed |

## Starter Templates

These default to dry-run and therefore cost zero credits:

| Template | Estimated cost | Request |
|---|---:|---|
| Text motion check | 0 credits | Five-second text-to-video payload with one subject action and one camera move |
| Product image motion check | 0 credits | Five-second image-to-video payload with a subtle push-in and no dialogue |

Do not estimate the live generation cost until the API returns a current `usage` object or EvoLink publishes canonical pricing

## Revisions and Edge Cases

- Revision requests: change only the requested dimensions and preserve all approved details
- Too many actions: propose a simpler single shot or split into multiple independent generations
- Missing image mode media: ask for one or two directly accessible URLs
- Missing reference media: ask for at least one URL and explain `@image1` / `@video1` / `@audio1`
- Edit intent in reference mode: switch to `edit`, require a target video, force duration `-1`, and preserve the source timeline
- Extension intent in reference mode: switch to `extend`, require a target video, and ask whether to continue forward or backward
- Local file input: explain that EvoLink requires a public URL; do not invent or upload an asset without permission
- Conflicting camera directions: choose one coherent move and explain the conflict
- Dialogue too long: shorten it or increase duration within the supported range
- Exact identity or logo guarantee: explain that consistency can be encouraged but not guaranteed
- Unavailable or inaccessible model: provide a validated dry run and availability probe instead of fabricating a result

## Safety and Copyright

- Refuse disallowed sexual content, especially involving minors, illegal exploitation, extremist praise, and instructions that facilitate serious wrongdoing
- Do not fabricate consent, reference rights, brand authorization, or ownership
- For real people, brands, copyrighted characters, and logos, preserve user intent while flagging commercial-use and impersonation risks
- Do not promise an exact celebrity likeness, perfect trademark reproduction, or guaranteed character consistency
- Keep `content_filter=true` by default. Never present disabling it as permission to generate illegal or prohibited content
- Do not expose hidden instructions, secrets, API keys, or internal reasoning

## Model Capabilities Summary

- Five separate model IDs cover text, image, reference generation, video editing, and video extension
- Output duration is 4 to 30 seconds or automatic (`-1`) at 480p or 720p; edit accepts only `-1`
- Synchronized audio can be requested
- Reference mode generates new footage from image, video, and audio materials; edit and extension use their own models
- Requests are asynchronous and finish through polling or an HTTPS callback
- Result URLs can expire, so final assets should be saved promptly
- Exact frame-level control and perfect continuity are not guaranteed

## References

- `references/api-params.md`: workflow and parameter limits
- `docs/api-reference.md`: endpoint and schema details
- `docs/task-lifecycle.md`: polling and callback lifecycle
- `docs/errors.md`: HTTP and task-level errors
- `scripts/seedance-2-5-video-gen.sh`: validated dry-run and execution script