Create controllable 2K video from text, frames, and image, video, or audio references.
MiniMax H3 is a general-purpose multimodal generation model that understands relationships across text, images, video, and audio. Through the AI Video API, developers can create 4–15 second clips at 768p or native 2K, use first and last frames, transfer motion from video, guide sound with audio references, and receive hosted results through one asynchronous REST workflow.
Model capabilities
Use MiniMax H3 from the same platform surface as the rest of your video and image stack: API keys, credits, logs, webhooks, and docs stay consistent across providers.
Describe how a prompt, character image, motion video, and voice or music reference should work together. H3 interprets the relationships as one creative instruction instead of forcing each medium into a separate generation tool.
MiniMax reports strong instruction following plus accurate presentation of text and brand information. That makes the model useful for commercial layouts where composition, messaging, product identity, and readable visual details need to stay aligned.
Attach up to three short reference videos so H3 can follow their motion, camera behavior, timing, or performance while rendering a new subject or visual direction described by the prompt and other references.
H3 jointly models video and sound and produces native stereo audio. Reference audio can guide voice, music, or sound intent when it is paired with at least one image or video reference, without a separate audio-generation request.
Use one image as the first frame or two images as the first and last frames. The dedicated frame workflow follows the input image ratio and gives teams a practical way to control scene entry, transition direction, and final composition.
Submit one standard video-generation request, receive a task ID immediately, then poll the task or use a callback URL. Completed MiniMax output is mirrored to platform storage and returned in the same output shape as other AI Video API models.
Write a prompt up to 7,000 characters and choose text-only, first/last-frame, or omni-reference generation. Add only public HTTP(S) image, MP4/MOV video, and MP3/WAV audio URLs that the upstream service can retrieve.
Set a 4–15 second duration and select 768p or 2K. Text-to-video uses a fixed ratio such as 16:9, frame mode follows the image, and multimodal reference generation can use adaptive aspect ratio.
POST to the unified video endpoint with model minimax-h3. The API validates reference limits, detects billable reference-video duration, pre-charges the estimate, and returns a task ID for callbacks or status polling.
When the task succeeds, the temporary MiniMax result is copied into platform storage before completion is published. Read the final URL from output.urls and retain the normal task metadata for logs and downstream automation.
AI Video APIcurl -X POST https://api.aivideoapi.ai/v1/videos/generations \
-H "Authorization: Bearer sk-your-api-key" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3",
"callback_url": "https://your-server.com/webhook",
"input": {
"prompt": "Use the reference camera motion while keeping the product identity and matching the reference voice",
"generation_type": "omni_reference",
"image_urls": ["https://example.com/product.png"],
"video_urls": ["https://example.com/camera-motion.mp4"],
"audio_urls": ["https://example.com/voice.mp3"],
"resolution": "2k",
"duration": 8,
"aspect_ratio": "adaptive"
}
}'Why choose MiniMax H3
MiniMax describes H3's instruction-following foundation as Contextual Omni Representation. Language connects the target clip to multiple context items and expresses relationships among those items, allowing open-ended directions such as combining a camera move, a character image, and a voice reference in one request.
The H3-VAE tokenizer was redesigned for reconstruction quality, learnability, and higher compression. MiniMax reports a fourfold effective sequence-length gain, reducing training and inference cost while providing the architectural foundation for native 2K generation rather than relying only on a conventional post-process upscale.
The H3-Omni Transformer is designed around task generalization and efficient multimodal workloads. MiniMax separates the understanding and generation workloads during training and reports that balancing their different compute patterns improved end-to-end training throughput by nearly 30 percent.
For 2K output, H3 regenerates its own lower-resolution result while reusing the original multimodal context. The model can revisit prompt and reference details during high-resolution rendering, which MiniMax says helps preserve small text and fine information that a standalone super-resolution module would otherwise need to infer.
Frequently asked questions
The MiniMax H3 API is AI Video API's asynchronous REST integration for MiniMax H3, a general-purpose multimodal video model. It accepts a required text prompt plus optional image, video, and audio context, and can also animate one or two frame images. Requests use the shared POST /v1/videos/generations endpoint with model minimax-h3, return a task ID immediately, and provide hosted video output through polling or callback completion.
Use omni_reference for text-to-video or any supported combination of reference images, videos, and audio. Use first_and_last_frames when image_urls contains a first frame and optional last frame; that mode cannot include video or audio references. Text-only generation defaults to 16:9 and cannot use adaptive ratio, while reference generation can use adaptive and frame mode follows the supplied image dimensions.
An omni-reference request supports up to nine image URLs, three video URLs, and three audio URLs. Each reference video must be 2–15 seconds and their combined duration cannot exceed 15 seconds. Audio cannot be the only reference medium, so include at least one image or video with audio_urls. Every reference must be a public HTTP(S) URL; data URIs and private MiniMax file identifiers are not accepted by this gateway.
At 768p, output video and input reference video both cost 22 credits per second. At 2K, both cost 34 credits per second. A five-second 2K text-to-video request therefore costs 170 credits. If the same request includes 12 seconds of reference video, the platform bills 17 total seconds at 34 credits, or 578 credits. The first five reference images are free, each additional image costs 9 credits, and reference audio is free.
Yes. MiniMax describes H3 as jointly generating video with native stereo sound rather than treating voice, sound effects, and music as isolated systems. In omni-reference mode, you can attach up to three MP3 or WAV audio references to guide voice, musical, or sound intent, provided the request also includes at least one reference image or video. There is no separate audio surcharge in the current AI Video API pricing formula.
Yes. MiniMax highlights V2V motion transfer as a core H3 capability. Provide a short public MP4 or MOV file in video_urls and describe which motion, camera move, rhythm, or performance should influence the result. You can combine that video with image references for subject or visual identity and with audio references for sound intent. The server measures all reference-video duration before pricing and validates the 15-second combined limit.
The API validates inputs, probes reference-video duration when needed, calculates the charge, and submits an asynchronous MiniMax task. The task moves from pending to processing and then completed or failed. You can poll GET /v1/tasks/{taskId} or provide callback_url. Successful results are copied from MiniMax's temporary delivery URL to platform storage before output.urls is published; failed generations receive a full credit refund.
Start with a clear creative instruction that names the role of every reference instead of listing disconnected assets. MiniMax positions H3 for advertising, branding, e-commerce, product design, UI/UX, gaming, film titles, and animated posters, with particular emphasis on instruction following, text and brand presentation, and controllable multimodal editing. Test a short 768p draft first when exploring direction, then move selected prompts and references to 2K for final delivery.
Pricing and usage
Applied to output duration plus detected reference-video seconds
Default resolution; applied to output and reference-video seconds
Each image after the fifth costs 9 credits
Up to three audio URLs when paired with image or video context
Create one API key, use one credit balance, and switch between video and image models without provider-specific plumbing.