Wan 3.0: Model Documentation & Integration Guide

Wan3.0 is an all-in-one video generation model from Wan-AI, hosted on DeepInfra. It covers reference-based generation, editing, replication, and driving in a single model, and it accepts images, video clips, audio, documents, and web links as input.

Use it for cinematic storytelling, social media content, and video editing. Every generation includes synchronized audio by default, and the model keeps characters stable and recognizable across a clip.

Key Capabilities

• Omni-modal reference: Parses documents, web pages, and complex images to drive video generation.
• Long clips with audio: Generates up to 30 seconds of video with a synchronized audio track by default.
• Character consistency: Keeps subjects stable and recognizable for the full duration of a video.
• Editing and extension: Handles style transfers such as clay art conversions, and extends existing videos forward or backward.

Technical Specifications and Benchmarks

Standard specifications

Quality benchmark

Artificial Analysis measures video quality through blind user votes in its Video Arena. Wan3.0 ranks first on both of its text-to-video leaderboards.

On the with-audio board, Wan3.0 leads Utopai X at 1,166 and Dreamina Seedance 2.5 at 1,159. The rank interval spans first to third, so the top models sit close together.

Comparison with Wan2.7

Both Wan2.7 models remain available, including Wan2.7 image-to-video. Alibaba introduced video editing in Wan2.7, and Wan3.0 carries it forward.

Getting Started

To use Wan3.0-Video, you need a DeepInfra account and an API key. DeepInfra applies a zero-retention policy to inputs and outputs.

Authentication

Create a key in the DeepInfra dashboard and pass it in the Authorization header of every request:

Authorization: Bearer <DEEPINFRA_API_KEY>

API endpoint basics

• Base URL: https://api.deepinfra.com/v1/inference
• Model path: /Wan-AI/Wan3.0-Video
• HTTP method: POST
• Content type: application/json

Making your first API call

The example below generates a short video with Python.

import os
import requests
 
url = "https://api.deepinfra.com/v1/inference/Wan-AI/Wan3.0-Video"
headers = {
    "Authorization": f"Bearer {os.environ['DEEPINFRA_API_KEY']}",
    "Content-Type": "application/json",
}
 
payload = {
    "prompt": "A cinematic drone shot flying over a misty pine forest at sunrise",
    "duration": 5,
    "resolution": "720P",
    "audio": True,
    "prompt_extend": True,
}
 
response = requests.post(url, headers=headers, json=payload)
print(response.json()["video_url"])
copy

To guide a generation with reference media, add a media array. Refer to each asset in the prompt as Image 1, Video 1, and so on, numbered in the order the assets appear in the array.

payload = {
    "prompt": "Image 1 walks through a rainy neon street at night",
    "media": [
        {"type": "reference_image", "url": "https://example.com/character.png"}
    ],
    "duration": 8,
}
copy

Configuration and Parameters

Each request needs either a prompt or media.

Common parameters

Media entries use one of these types: first_frame, last_frame, reference_image, reference_video, reference_audio, file, or link. Reference types cannot be combined with first_frame or last_frame. When a request includes a reference_video, the input and output durations together must stay within 30 seconds.

Input limits

Response Format

The API returns a JSON payload with a link to the generated video. An inference_status object reports the status, runtime, and cost of the request.

{
  "video_url": "https://example.com/video.mp4",
  "request_id": null,
  "inference_status": {
    "status": "succeeded",
    "runtime_ms": 0,
    "cost": 0.0
  }
}
copy

Pricing

Wan3.0-Video is billed at $0.20 per second of generated video. A 10-second clip costs $2.00, and a 30-second clip costs $6.00. DeepInfra lists the rate per second of generated output.

Conclusion

Wan3.0-Video combines omni-modal input with audio and video output in one model. Creators can turn text, images, and documents into finished video without assembling a multi-model pipeline.

• Range: Generates up to 30 seconds of video with integrated audio and multi-modal references.
• Control: Offers prompt_extend, adaptive aspect ratios, seeds, and optional watermarks.
• Simple billing: Charges a single per-second rate for generated video.

DeepInfra also hosts other video models that accept reference inputs, such as ByteDance Seedance 2.0, so you can compare outputs through the same API.