Vidu Q4 Preview Reference-to-Video Spicy

Vidu Q4 Preview Reference-to-Video Spicy (vidu/q4-preview-global) is Vidu's next-generation multimodal video generation model, offering high-consistency multi-subject reference generation alongside flexible first-frame and first-to-last frame image-to-video capabilities through a unified endpoint. It supports up to 15 reference images and up to 2 reference audio tracks (sounds) in reference mode, or 1–2 frame images in image mode. Capable of generating 4 to 16-second videos at resolutions ranging from 540p up to 4K (540p, 720p, 1080p, 2K, 4K) with native audio synthesis, it delivers cinematic motion, accurate subject consistency, and synchronized audio-visual storytelling across mainstream aspect ratios (16:9, 9:16, 3:4, 4:3, 1:1).

Base URL

https://api.icreat.ai

Authentication

All API requests must be authenticated with an API Key. You can obtain an API Key from the console.

export ICREAT_API_KEY="your-api-key-here"

HTTP Request Headers

import os

API_KEY = os.environ.get("ICREAT_API_KEY")
headers = {
    "Content-Type": "application/json",
    "Authorization": "Bearer " + API_KEY,
}

Protect your API Key Never expose your API Key in client-side code or public repositories. Use environment variables or a backend proxy.

Code Examples

Image and video generation uses a two-step async flow: submit a task to get task_id, then poll via query task result; the response includes status and result ([] while processing; on SUCCEEDED, result holds resources and costUSD is present). The examples below use the same task_id across both steps.

1. Submit Task

Send a generation request to the submit endpoint using a flat top-level JSON body. Both Reference-to-Video (mode: "reference") and Image-to-Video (mode: "image") share the same endpoint URL.

2. Query Task Result (Poll)

Use the task_id from submit to poll progress (repeat until terminal). The response includes status and result: result is [] while processing; on SUCCEEDED, result holds resources and costUSD is included; FAILED means the task failed.

Input Schema

Submit Task — Input

The request body is a flat top-level JSON object (do not wrap fields in input or parameters).

Total: 10 (prompt is also required when mode is "reference") Optional: 9

Query Task Result — Input

Total: 1 Required: 1 Optional: 0

Output Schema

Submit Task — Output

Total: 1

Query Task Result — Output

Total: variable

# vidu/q4-preview-global

> Vidu Q4 Preview Reference-to-Video Spicy (`vidu/q4-preview-global`) is Vidu's next-generation video generation model supporting multi-image/audio Reference-to-Video (`mode: "reference"`) and first-frame / first-to-last frame Image-to-Video (`mode: "image"`) via a single unified endpoint.

## Overview

Use the iCreat two-step async task API: submit a generation request with a flat JSON body, then poll the query task result endpoint; on success read resources from `result` (includes `costUSD`).

## API Info

- **Base URL**：`https://api.icreat.ai`
- **Submit endpoint (POST)**：`/v1/task/submit/vidu/q4-preview-global`
- **Query result endpoint (POST)**：`/v1/task/result`
- **Model ID**：`vidu/q4-preview-global`
- **Auth**：`Authorization: Bearer ${ICREAT_API_KEY}`

## Call Flow

1. **Submit**: POST submit path with a flat JSON body per Input Notes; response `{ "task_id": "..." }`
2. **Query result**: POST `/v1/task/result` with `{ "task_id": "..." }`; response includes `status` and `result` (`[]` while processing); on `SUCCEEDED`, `result` holds resources and `costUSD` is present; read `url` or `download_url` when `type` is `Video`

### Input Example (Reference-to-Video)

```json
{
  "mode": "reference",
  "images": [
    "https://example.com/ref-scene.jpg](https://example.com/ref-scene.jpg"
  ],
  "sounds": [
    "https://example.com/ref-voice.mp3](https://example.com/ref-voice.mp3"
  ],
  "prompt": "Generate a 5-second short clip based on the reference image and reference audio.",
  "duration": 5,
  "resolution": "720p",
  "aspectRatio": "16:9",
  "audio": true,
  "seed": 42
}

Input Notes

• Top-level body: Flat JSON object (do NOT wrap in input or parameters).
• mode (optional): "image" (default) or "reference" (case-insensitive).
• images (required): Array of image HTTP/HTTPS URLs. 1–15 images when mode="reference"; 1–2 images when mode="image" (1 for first-frame, 2 for first-to-last frame).
• prompt (conditional): Required and non-empty when mode="reference"; optional when mode="image". Max 2000 Unicode characters.
• sounds (optional): Array of audio HTTP/HTTPS URLs (max 2). Supported ONLY when mode="reference". Passing non-empty sounds in image mode causes a 400 error.
• duration (optional): Integer 4–16 seconds (default 5).
• resolution (optional): "540p", "720p" (default), "1080p", "2K", or "4K" (case-insensitive).
• aspectRatio (optional): "16:9" (default), "9:16", "3:4", "4:3", or "1:1". Effective ONLY when mode="reference"; do NOT pass when mode="image".
• audio (optional): Boolean (default true). Set to false to mute audio output.
• isRec (optional): Boolean. AI prompt recommendation, supported ONLY when mode="image"; do NOT pass when mode="reference".
• seed (optional): Integer seed for reproducibility.

Output Notes

• Poll: read status; result is [] while processing (SUBMITTED or IN_PROGRESS)
• Success: status is SUCCEEDED, result is [{ "type": "Video", "url": "...", "download_url": "..." }] plus costUSD

Notes

• Use the same task_id across both steps; result is [] while processing — keep polling
• FAILED and NOT_FOUND are terminal — do not keep polling; check request parameters or media URLs
• Generated media URLs are temporary; download and persist the video promptly rather than hardcoding the temporary URL