Seedance 2 Video Gen

Werydance 2.0 video generator (almighty-ready)

Generate Werydance 2.0 videos on WeryAI with one skill for text-to-video, image-to-video, first-frame/last-frame transition video, multi-image storyboard video, and almighty mixed-reference video generation. Animate and transform scenes from single images, storyboard frames, or mixed references while keeping subject and motion continuity.

Dependencies: scripts/video_gen.js at the skill package root (alongside SKILL.md) + WERYAI_API_KEY + Node.js 18+. No other Cursor skills are required.

• text-only generation
• single-image animation
• multi-image storyboard generation
• first-frame/last-frame compatibility aliases
• almighty mixed references (images + videos + audios + prompt)

Runtime secret: only WERYAI_API_KEY.

Core upgrade

This skill now supports four input modalities for richer control:

• Image reference: composition, character, and detail lock
• Video reference: camera language, action rhythm, effect imitation
• Audio reference: mood and timing guidance
• Text prompt: intent and constraints

Smart routing policy:

• If payload includes videos or audios, route to POST /v1/generation/almighty-reference-to-video.
• Otherwise keep legacy routing (text-to-video, image-to-video, multi-image-to-video).

Prerequisites

• WERYAI_API_KEY must be set.
• Node.js 18+.
• Prefer public https URLs for media.
• Local files are supported and auto-uploaded by this skill via POST /v1/generation/upload-file.

export WERYAI_API_KEY="your_api_key_here"

Quick checks

node scripts/video_gen.js models --mode text_to_video
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"A glowing koi swims through ink clouds","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}' --dry-run

Modes and input shapes

Compatibility aliases still supported:

• first_frame + last_frame -> normalized to ordered images
• image + last_image -> normalized to ordered images

Almighty constraints

Almighty validations in this skill:

• prompt max 500 chars
• images <= 9
• videos <= 3
• audios <= 3
• mixed file total (images + videos + audios) <= 12
• at least one of images or videos
• duration must be 5-15 seconds
• video_number must be 1-4

Notes:

• Backend enforces media-duration constraints for reference clips.
• Per current OpenAPI contract, almighty duration floor is 5s.

Local upload behavior

Supported local files before generation submit:

• Images: jpg/jpeg/png/webp/gif, max 10MB
• Videos: mp4/mov, max 50MB
• Audios: mp3/wav, max 50MB

Upload endpoint used by this skill:

• Shot scale, angle, and lens feel.
• Camera movement or a clear lock-off choice.
• Lighting direction, material response, and overall mood.
• Subject motion paced to the selected duration.
• One clear visual payoff in the final shot.
• Output framing such as 9:16 vertical when using the default aspect ratio.
• Audio (default-on): Add a labeled Audio: block (ambience, layered SFX, subtle foley—generic, non-copyrighted) even if the user never mentioned sound. JSON generate_audio defaults to true for models that support audio; use false and omit Audio: only when the user explicitly wants silent output.

Dry-run behavior:

• --dry-run never uploads local files
• output includes uploadPreview to show which local paths would be uploaded

Prompt workflow (mandatory)

Before paid submit, expand brief into production prompt:

1. Decide mode.
2. Map each media to explicit role.
3. Build 2-4 timeline beats by duration.
4. Write one action arc + one final payoff.
5. Add minimal constraints for continuity.

Always show full prompt and all params before submit, then wait for explicit confirmation.

@asset layered referencing

Use explicit role mapping in prompt drafts:

• Visual anchor: @图片1 / @图1
• Camera/motion style: 参考@视频1运镜
• Action rhythm: 参考@视频2动作节奏
• Music/sound bed: @音频1用于配乐

Good:

• @图1为首帧，参考@视频1的跑步动作，镜头从正面切到侧面，@音频1做节奏底。

Bad:

• 用这些素材生成一个视频。

High-value templates

Video extend

• 将@视频1延长 5s，新增段落为...，保持人物与场景连续。
• Set duration to the added segment length.

Video fusion

1. Decide which path fits the request: text, image, multi-image, or first-frame/last-frame.
2. Collect the user's brief, duration, optional aspect_ratio, optional resolution, generate_audio (default true unless the user wants silent), and reference image URLs if the mode uses assets (prefer public https).
3. Expand the prompt with the rules from ## Prompt expansion (mandatory). If the request needs tighter control, read references/seedance-prompt-optimization.md and explicitly decide mode, asset mapping, timeline beats, and negative constraints before writing the final prompt.
4. If you are unsure about current multi-image support or field limits, run node scripts/video_gen.js models --mode multi_image_to_video before a paid request.
5. Show a confirmation table with the full expanded prompt, model, duration, aspect ratio, resolution, audio choice, and any image URLs.
6. After explicit confirmation, run node scripts/video_gen.js wait --json '...'.
7. Return the JSON result. On success, surface the video URLs. On failure, report errorCode and errorMessage and suggest the likely parameter fix.

CLI reference

# Check live model metadata
node scripts/video_gen.js models
node scripts/video_gen.js models --mode text_to_video
node scripts/video_gen.js models --mode image_to_video
node scripts/video_gen.js models --mode multi_image_to_video

# Text-to-video (non-empty model required; default audio on)
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"A paper crane unfolds into a real bird, cinematic lighting, 9:16 vertical. Audio: soft paper rustle, gentle whoosh, quiet room tone.","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}'

# Single image to video
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"Animate this portrait with subtle hair and cloth movement. Audio: faint fabric rustle, subtle room ambience.","image":"https://example.com/frame.png","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}'

# Multi-image to video
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"Turn these storyboard frames into one coherent reveal shot. Audio: light transition whooshes, soft underscore.","images":["https://example.com/1.png","https://example.com/2.png","https://example.com/3.png"],"duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}'

# First-frame / last-frame guided video
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"Start on the first frame and transition naturally to the last frame with the same subject and environment. Audio: smooth morph tone, airy ambience.","first_frame":"https://example.com/start.png","last_frame":"https://example.com/end.png","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}'

# WaveSpeed-style compatibility alias
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"Transition smoothly from the start image to the end image. Audio: soft blend whoosh, quiet room.","image":"https://example.com/start.png","last_image":"https://example.com/end.png","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}'

# Preview without spending credits
node scripts/video_gen.js wait --json '{"model":"WERYDANCE_2_0","prompt":"A glowing koi swims through ink clouds. Audio: water shimmer, subtle chime.","duration":5,"aspect_ratio":"9:16","resolution":"720p","generate_audio":true}' --dry-run

# Poll an existing task
node scripts/video_gen.js status --task-id <task-id>

Poll timeout classes

• task_class=short: default 5 minutes (300000ms), intended for text-to-video.
• task_class=long: default 20 minutes (1200000ms), intended for image/multi-image/almighty.
• task_class=auto (default): infer from effective mode.
• WERYAI_POLL_TIMEOUT_MS remains the highest-priority explicit override.

Compliance notice

• Do not store WERYAI_API_KEY in repository files.
• Platform compliance for realistic human face references is handled by backend policy; this skill does not implement client-side blocking.

Definition of done

Done when the user receives at least one playable video URL, or a clear failure with errorCode / errorMessage and a concrete next step. The submitted payload must stay within the selected model's supported field set. With default generate_audio: true, the prompt should include Audio: prose unless the user requested silent generation.

Boundaries (out of scope)

• Do not use local file paths for references unless you have reviewed the assembled scripts/video_gen.js and explicitly consent to local read-and-upload to WeryAI; otherwise prefer public https URLs.
• Do not invent undocumented API fields beyond the JSON surface in this skill.
• Do not assume multi-image or first/last-frame support without checking live metadata when the behavior matters.
• Do not store the secret value of WERYAI_API_KEY inside the repository.
• Do not copy parameter names blindly from another provider without checking the normalized fields in this document.