MiniMax H3 Image-to-Video API

MiniMax H3 is a general-purpose, omni-modal video generation model released by MiniMax in July 2026. It generates high-quality videos from multimodal content (text, images, video, and audio). This document covers its Image-to-Video capability: generating video from first-frame/last-frame images plus a text prompt, with 2K direct output, first/last-frame control, and multimodal reference-based video generation, served through an out-of-the-box REST inference API.

Model Positioning

MiniMax H3 Image-to-Video targets video creation workflows that start from an image. It is the image-to-video endpoint of the H3 omni-modal generation system. Unlike the Text-to-Video endpoint, requests must include at least one reference image (first frame, or first and last frames) along with a text prompt describing camera movement and visual changes. The model animates the image while preserving subject consistency, making it ideal for tasks with strict start/end requirements such as product display, character transitions, and first/last-frame interpolation.

Core Capabilities

2K direct output

Native 2K (2160P) resolution output, not upscaled; a 768P tier is also available to control cost. Both tiers are billed per output duration.

First/last-frame control

Set the role field of image_url elements in the content array to first_frame and last_frame to precisely control the start and end frames of the shot. Providing only a first frame enables first-frame-to-video mode.

Multimodal reference input

content supports four element types — text / image_url / video_url / audio_url — enabling mixed reference materials combined with a text prompt for style transfer and content consistency.

Native stereo audio

24 FPS cinematic frame rate with 32 kHz stereo audio generated in sync with the visuals, no separate dubbing needed.

Asynchronous task queue

Standard three-step flow: submit task → query status → get result, suitable for long-running jobs and batch production.

Pricing

Total = unit price × output video duration. Billed by actual generated duration with no minimum-length threshold.

Use Cases

• E-commerce product animation: turn product images into showcase videos with clean start/end frames
• Advertising and brand storyboards: animate static storyboard frames into dynamic previews
• Short-video and social media assets: batch image-to-video production with controlled per-clip cost
• Character transitions and camera transitions: precise first/last-frame specification for smooth cuts
• Pre-visualization: animate storyboard frames to help directors and art teams communicate

Model Comparison

MiniMax H3 Image-to-Video vs. H3 Text-to-Video

MiniMax H3 Image-to-Video vs. Hailuo 2.3 and Hailuo 02

Note: Hailuo 2.3 is billed per video (domestic list price from 2 CNY per clip), while the H3 series is billed per second; convert by total duration when comparing unit prices directly.

Why Choose MiniMax H3 Image-to-Video?

• 2K direct output + native stereo audio: visuals and sound generated in one pass, no upscaling or dubbing
• Precise first/last-frame control: stable results for shots with strict start/end requirements (transitions, product display)
• Multimodal reference: mixed image/video/audio input with strong complex-instruction following
• Transparent per-second billing: unit price × duration, predictable cost with no hidden fees
• Open ecosystem: built on 33B open-sourced weights, deployable locally or customizable

API

This interface is an asynchronous task-based REST API. Submit endpoint: https://api.icreat.ai/v1/task/submit/minimax/h3-video/image-to-video. The flow has three steps: POST the task parameters to the submit endpoint and receive a task_id; poll the query endpoint with the task_id until the status is SUCCEEDED; then call the result endpoint to fetch the generated video. A callback_url can also be provided to receive task status-change notifications.

Key request parameters: content (multimodal input array; must include a non-empty text item as the prompt), resolution (768P or 2K), duration (integer 4–15), ratio (aspect ratio, default adaptive), callback_url (optional callback address), and aigc_watermark (optional AIGC watermark toggle, default false).

Specifications

Architecture

MiniMax H3 Image-to-Video uses an asynchronous task-queue architecture. Requests enter the task queue via the submit endpoint and are processed by the H3 omni-modal generation system (Context-IR instruction understanding + H3-VAE + H3-Omni Transformer + In-context Regeneration). The system first distills the input instructions into a context intermediate representation, then generates a video frame sequence with native stereo audio guided by the reference materials, finally outputting at 2K or 768P. Task status and results are queried through dedicated endpoints, naturally supporting high-concurrency batch calls.

Notes

• duration must be an integer between 4 and 15; out-of-range values return an error
• content must contain a non-empty text item (prompt is required); images alone cannot be submitted
• When ratio is not specified it defaults to adaptive, letting the system pick the most suitable aspect ratio
• In first/last-frame mode, image_url elements must mark first_frame / last_frame via role
• The 2K tier costs more per second than 768P; estimate cost in advance for long videos (total = unit price × duration)
• Fetch results via the result endpoint promptly after completion; result retention period is subject to the platform

FAQ

What is the difference between MiniMax H3 Image-to-Video and Text-to-Video?

Both belong to the H3 series and share the same resolution (768P/2K), duration (4–15 seconds), and asynchronous task flow. The difference is input: Image-to-Video requires an image (first or first/last frame) and animates it, with precise first/last-frame control; Text-to-Video takes only a text prompt for creative generation without reference visuals.

How do I specify first and last frames in a request?

Add two image_url elements to the content array with role set to first_frame and last_frame respectively. If only a first frame is provided, the model runs in first-frame-to-video mode and generates the last frame automatically.

Why must the submission example include image input?

Because this endpoint is image-to-video. A request with only a text prompt and no image is a text-to-video task and should use the text-to-video endpoint instead.

How is billing calculated?

Billed by output duration: total = unit price × output video duration. 768P is $0.10/sec and 2K is $0.14/sec; a 5-second clip costs $0.50 and $0.70 respectively.