Vidu Skills

Vidu Video and Image Generation Skill (Vidu 音视频/图像生成技能)

Generate AI videos and images with Vidu (生数) via direct API calls — text-to-image, text-to-video, image-to-video, start-end frame, reference-based generation, and material elements, up to 1080p/2K/4K. Use curl with VIDU_TOKEN.

Execution model: use curl (direct API)

All execution is done by calling the official Vidu API with curl (or any HTTP client). Base URL: $VIDU_BASE_URL (default https://service.vidu.cn for mainland China; https://service.vidu.com for overseas).

Required headers for all requests:

Main endpoints:

• Create upload: POST $VIDU_BASE_URL/tools/v1/files/uploads → get put_url, id
• PUT image: PUT raw image bytes to put_url → get ETag
• Finish upload: PUT $VIDU_BASE_URL/tools/v1/files/uploads/{id}/finish → get ssupload:?id={id}
• Submit task: POST $VIDU_BASE_URL/vidu/v1/tasks → get task_id (response id)
• Get task result: GET $VIDU_BASE_URL/vidu/v1/tasks/{task_id} → get state, creations[].nomark_uri
• Task state (SSE): GET $VIDU_BASE_URL/vidu/v1/tasks/state?id={task_id} with Accept: text/event-stream — return SSE stream to the user; do not wait for terminal state. Events include state, estimated_time_left, err_code, queue_wait_time (排队预测时间, unit: minutes).
• Pre-process reference: POST $VIDU_BASE_URL/vidu/v1/material/elements/pre-process
• Create reference: POST $VIDU_BASE_URL/vidu/v1/material/elements
• List elements: GET $VIDU_BASE_URL/vidu/v1/material/elements/personal

Key Capabilities

• text-to-image (文生图) — POST /vidu/v1/tasks with type: "text2image", input.prompts (text only), settings. Resolution 1080p, 2K, 4K.
• text-to-video (文生视频) — POST /vidu/v1/tasks with type: "text2video", input.prompts (text only), settings.
• image-to-video (图生视频) — Upload one image (Create upload → PUT → Finish) to get ssupload:?id=...; then POST /vidu/v1/tasks with type: "img2video", prompts (text + image).
• head-tail-image-to-video (首尾帧生视频) — Upload two images; POST /vidu/v1/tasks with type: "headtailimg2video", prompts (text + image1 + image2).
• reference-to-image (参考生图) — Image(s) + reference(s) + text (text required; image + reference combined at most 7). POST /vidu/v1/tasks with type: "reference2image"; Q2 only, do not send transition, duration is 0.
• reference-to-video (参考生视频) — Image(s) + reference(s) + text (text required; image + reference combined at most 7). POST /vidu/v1/tasks with type: "character2video"; Q3 or Q2, do not send transition.
• Create References (创建主体) — POST pre-process → POST material/elements (images must be uploaded first). Query list: GET /vidu/v1/material/elements/personal.
• Query task (查询任务) — GET /vidu/v1/tasks/{task_id} for result; or GET /vidu/v1/tasks/state?id={task_id} for SSE stream.

Setup

1. Obtain a VIDU token (e.g. from the official Vidu console).
2. Set environment variables:
   • export VIDU_TOKEN="your-token" (required / 必填)
   • export VIDU_BASE_URL=https://service.vidu.cn (mainland China / 国内, default / 默认) or https://service.vidu.com (overseas / 海外)
3. Dependency: curl or any HTTP client that can send JSON and binary PUT. No Python or scripts required for execution.

Data usage and privacy note

IMPORTANT: This skill sends user-provided data to Vidu’s servers:

• Text prompts → Vidu API
• Image bytes (uploaded files) → Vidu API servers (service.vidu.cn or service.vidu.com)
• Task parameters (settings, model version, etc.)

Before using this skill, confirm that sending your content to Vidu is acceptable for your privacy and intellectual property requirements. Data handling follows Vidu’s official policy.

Security recommendations:

• Create a token with limited scope if possible
• Avoid using production/privileged tokens for initial testing
• Review Vidu’s terms of service and privacy policy

Vidu Terms & Privacy:

• Overseas: https://www.vidu.com/terms
• Mainland China: https://www.vidu.cn/terms

Overview

Vidu media generation is asynchronous: submit a task → get task_id → use task_id to query status/result (GET /vidu/v1/tasks/{task_id} or SSE /vidu/v1/tasks/state?id=) when needed.

• text-to-image (文生图): Text only. Duration 0, model_version 3.1. aspect_ratio optional. resolution 1080p, 2K, 4K (default 2K).
• text-to-video (文生视频): Text only. Q3 duration 1–16, aspect ratios 16:9/9:16/1:1/4:3/3:4, transition pro/speed; Q2 duration 2–8, do not send transition.
• image-to-video (图生视频): One image + one text. Aspect ratio from input image (do not send aspect_ratio). Q3 duration 1–16, Q2 duration 2–8, transition pro/speed.
• head-tail-image-to-video (首尾帧生视频): Two images (start frame, end frame) + one text. Q3 1–16s, Q2 2–8s, transition pro/speed.
• reference-to-image (参考生图): Image + reference + text (combinations); text required. Image + reference at most 7, at least one. Q2 only, duration 0, reference via type: "material".
• reference-to-video (参考生视频): Image + reference + text (combinations); text required. Image + reference at most 7, at least one. Q3 duration 1–16, Q2 duration 2–8. Do not send transition. References in prompts via type: "material", material.id, material.version.
• Create References (创建主体): Upload 1–3 images, name and optional description; must call POST /vidu/v1/material/elements/pre-process first, then POST /vidu/v1/material/elements. Use pre-process recaption when description is omitted. Response includes element id and version for reference-to-video.
• Search References (查询主体): GET /vidu/v1/material/elements/personal with pager.page, pager.pagesz, keyword, modalities; returns elements[].id, version.

See Supported task list below and references/parameters.md.

Supported Task List

When building the POST /vidu/v1/tasks body, ensure the user’s request matches one of the supported task types and constraints below. All parameters are passed in the request body (type, input.prompts, settings). references/parameters.md has the same list for quick lookup.

Model version

• Q3 → model_version: "3.2"
• Q2 → model_version: "3.1"

• text-to-image (文生图): Text only; set model_version to 3.1, duration to 0, resolution defaults to 2K (1080p/2K/4K).
• text-to-video (文生视频): Text only; do not send transition for Q2.
• image-to-video (图生视频): Exactly 1 image + 1 text; do not send aspect_ratio in settings.
• head-tail-image-to-video (首尾帧生视频): Exactly 2 images (start, end) + 1 text; order is start frame then end frame.
• reference-to-image (参考生图): Image + reference + text; Q2 only, duration 0, resolution defaults to 2K, do not send transition.
• reference-to-video (参考生视频): Image + reference + text (text required; image + reference combined at most 7); Q3 or Q2; 不要传 transition.

Create References

Create a material element for use in character2video or reference2image: upload 1–3 images, name and description; API returns element id and version. Flow: First call POST $VIDU_BASE_URL/vidu/v1/material/elements/pre-process (required even if user provides description; body: components, name, type: "user"). Response includes recaption (style, description). Then call POST $VIDU_BASE_URL/vidu/v1/material/elements with the pre-process response id, same components, and recaption (or user description).

Pre-process: POST $VIDU_BASE_URL/vidu/v1/material/elements/pre-process. See references/api_reference.md §3b.

Create element: POST $VIDU_BASE_URL/vidu/v1/material/elements. Body: id (from pre-process), name, modality: "image", type: "user", components (1–3 items: content = ssupload:?id={id}, src_img = ssupload:?id={id}, content_type: "image"), version: "0", recaption. See references/api_reference.md §3c.

Prerequisite: Each image must be uploaded first (Create upload → PUT to put_url → Finish) to get ssupload:?id={id}; use these in components.