Explore/MiniMax H3 Highspeed
Video

MiniMax H3 Highspeed

Create 4–15 second videos with MiniMax H3 Highspeed, guided by image, video and audio references. Combine up to 9 images, 3 video clips and 3 audio clips at 480P or 768P.

View docs
import { config, higgsfield } from "@higgsfield/client/v2";config({  credentials: process.env.HF_CREDENTIALS,});const result = await higgsfield.subscribe(  "minimax/h3-highspeed/reference-to-video",  {    input: {      "prompt": "A cinematic scene at sunset",      "duration": 5,      "audio_urls": [],      "image_urls": [        "https://example.com/reference.png"      ],      "resolution": "768P",      "video_urls": [],      "aspect_ratio": "auto",      "aigc_watermark": false    },    withPolling: true,  },);console.log(result);

MiniMax H3 Highspeed uses the model ID minimax/h3-highspeed/reference-to-video. Create 4–15 second videos with MiniMax H3 Highspeed, guided by image, video and audio references. Combine up to 9 images, 3 video clips and 3 audio clips at 480P or 768P.

1. Calling the API

Send requests to https://api.higgsfield.ai/minimax/h3-highspeed/reference-to-video with a JSON body matching the parameters below.

Install

Install an official server-side SDK. cURL needs no package.

npm install @higgsfield/client

Setup

Keep credentials in server-side environment variables. The SDKs accept the same KEY_ID:KEY_SECRET value under their documented variable names.

export HF_CREDENTIALS="YOUR_KEY_ID:YOUR_KEY_SECRET" # TypeScript export HF_KEY="YOUR_KEY_ID:YOUR_KEY_SECRET" # Python and cURL

Request parameters

ParameterTypeRequiredDefaultDescription
promptstringYes—Prompt. Minimum length: 1. Maximum length: 7000.
durationintegerNo5Duration. Minimum: 4. Maximum: 15.
audio_urlsarray[string]Conditional[]Reference audio.
image_urlsarray[string]Conditional[]Reference images.
resolutionstringNo"768P"Resolution. Options: 480P, 768P.
video_urlsarray[string]Conditional[]Reference videos.
aspect_ratiostringNo"auto"Aspect ratio. Options: auto, adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, 9:16.
aigc_watermarkbooleanNofalse

2. Authentication

The official SDKs read credentials from the server environment and send the required Authorization: Key KEY_ID:KEY_SECRET header.

API Key

Never expose the secret in browser-side code or commit it to source control. The TypeScript SDK is server-side only. For direct HTTP calls, use:

Authorization: Key $HF_KEY

3. Queue

subscribe submits the asynchronous request and waits for a terminal result. The TypeScript SDK polls automatically when withPolling: true; the Python SDK's synchronous subscribe call also waits. A completed request returns the generated file in the video field.

Submit and wait

Replace placeholder media URLs with accessible URLs for your own files.

import { config, higgsfield } from "@higgsfield/client/v2"; config({ credentials: process.env.HF_CREDENTIALS, baseURL: "https://api.higgsfield.ai", }); const result = await higgsfield.subscribe( "minimax/h3-highspeed/reference-to-video", { input: { "prompt": "Create a gentle camera move that preserves the subject and visual style of the reference", "duration": 5, "resolution": "768P", "aspect_ratio": "16:9", "image_urls": [ "https://example.com/reference.png" ] }, withPolling: true, }, ); console.log(result);

Custom polling

For direct HTTP integration, submit the request and poll the returned status_url using the same API credentials until it reaches a terminal status. Use the returned cancel_url to request cancellation while the request is still queued. See the request lifecycle guide for response states.

4. Using this model

Provide a non-empty prompt of at most 7,000 characters. duration is an integer from 4 to 15 seconds, including every whole-second value in that range; it defaults to 5. resolution accepts 480P or 768P and defaults to 768P.

Provide at least one non-empty reference list: image_urls (up to 9), video_urls (up to 3), or audio_urls (up to 3). These references guide the generation; they are not first or last keyframes. Use the Image to Video variant for keyframe control.

Reference media is prepared before generation. Video clips may be resized, transcoded, trimmed or extended; the prepared video references have a combined duration of at most 14.8 seconds. Input-video charges use that prepared duration, which can differ from the original files.

Use the aspect_ratio values listed in the parameter table. Where offered, auto and adaptive derive framing from the input; text-only requests fall back to 16:9. aigc_watermark is fixed to false.

Use URLs accessible to the service. The documentation examples use placeholder media URLs; replace them with your own image or media URL before sending a request.

5. Pricing

ResolutionOutput video per secondInput video per second
480P$0.04$0.04
768P$0.06$0.06

The total price is (output seconds + prepared input-video seconds) × the rate for the selected resolution, plus $0.03 for each reference image beyond the first 5. The first 5 reference images are included, and audio references have no separate charge. Input-video duration is measured during preprocessing, so the final amount depends on the prepared media. For example, a 5-second output with 3 seconds of prepared video references and 6 images at 768P costs $0.51.