MiniMax H3 for AI persona video: references and sound

MiniMax H3 is MiniMax’s video model from 31 July 2026. It makes clips of 4 to 15 seconds at up to 2K, with stereo sound in the same run. For an AI persona?AI persona An AI persona is an invented character. Her images, videos and captions are generated with AI instead of photographed, but she is presented and followed like any other account. Open the glossary, use reference-to-video with up to nine images of her. Use MiniMax’s own API: the open-weights licence excludes the EU, UK, US and South Korea.

What MiniMax H3 is and what it makes

MiniMax H3 turns text, images, video and audio into a short clip with sound. MiniMax sells it through its API, and its video app Hailuo advertises it on the homepage.

Jul. 31, 2026, MiniMax H3: “A new-generation open general-purpose multimodal video model. It understands creative intent across multimodal context — text, image, video, and audio — and delivers more natural, coherent generation and expression.”

MiniMax API docs — Release notes (opens in a new tab)checked 5 Oct 2026

“MiniMax-H3: supports text-to-video, image-to-video (first / last frame), and reference-to-video; 768P / 2K resolution, 4–15s duration.” “MiniMax-H3-Max: the fast generation variant. Supports text-to-video, image-to-video (first / last frame), and reference-to-video (reference image / video / audio); 480P / 768P resolution, 2K is not supported; 5–15s duration.” Ratios include 9:16.

MiniMax API docs — Create Video Generation Task (opens in a new tab)checked 5 Oct 2026

H3 “can generate video with native stereo audio at resolutions up to 2K and durations of up to 15 seconds.” Output frame rate 24 FPS, output audio “32 kHz stereo”. “Stable support for 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, and Spanish.”

Hugging Face — MiniMax-H3 model card (opens in a new tab)checked 5 Oct 2026

Page title: “Hailuo AI: MiniMax H3 LIVE NOW. Top-Tier Quality, Versatile References”.

Hailuo AI — homepage (opens in a new tab)checked 5 Oct 2026

For a persona account, three details matter. Clips are short, so a Reel is several clips cut together. The format 9:16 is supported. And every clip comes with sound, whether you want it or not.

MiniMax H3 reference-to-video: how to keep your persona’s face

Use reference-to-video and upload your best images of her as references. Image-to-video starts from one frame. Reference-to-video lets her appear in any scene you describe, measured against several images of her face.

Reference limits: “Images: ≤ 9; width/height in [256, 5760]”, “Videos: ≤ 3 clips; per-clip duration [2, 15] s; total duration ≤ 15 s”, “Mixed input is capped at 12 files in total.”

MiniMax API docs — Video generation guide (opens in a new tab)checked 5 Oct 2026

The API reference adds one rule: “Image-to-video and reference-to-video are mutually exclusive.” So you either give H3 a first frame, or you give it a set of references. Not both.

MiniMax’s API docs do not say how to point at one image in the prompt. WaveSpeed’s page for its own H3 version does:

“Refer to every input in the prompt with an angle-bracket tag […]. The tags must be written exactly like that, with the brackets — plain text such as Picture 1 is treated as ordinary words, not a reference.” “Assign a job to each reference. Explicit assignments work far better than leaving the model to guess.” “A reference video’s own soundtrack is used automatically”. Images are numbered Picture 1 to Picture 9 in the order you upload them, videos Video 1 to Video 3, audio clips Audio 1 and up.

WaveSpeedAI — MiniMax H3 Reference-to-Video (open weights) (opens in a new tab)checked 5 Oct 2026

The tags look like this:

<Picture 1>  <Picture 2>  <Video 1>  <Audio 1>

A routine for her face:

1. Pick five images of your persona. Front view, both three-quarter views, one full-body image and one in different light. Five is the number that is free on MiniMax’s API. The character consistency guide explains how to build that set.
2. Use only images of your own generated persona. Never a photo of a real person. MiniMax’s terms make you guarantee you own what you upload, more on that below.
3. Give every image one job in the prompt. Say that pictures 1 to 5 show the same woman and are only for her identity.
4. Describe one action and one camera move. A 5-second clip has room for one thing.
5. Check the last frame against picture 1. Drift?Drift Drift is the slow change in a character’s face over time, until images made weeks apart no longer look like the same person. Open the glossary shows at the end of a clip. If the eyes, nose or jaw moved, generate again.

A prompt for WaveSpeed’s tag format:

<Picture 1>, <Picture 2>, <Picture 3>, <Picture 4> and <Picture 5> show the same woman, Mara,
in her late twenties. Use them only for her face, hair and body proportions.
Mara walks along a sunny harbour promenade towards the camera, smiles and pushes her hair
behind her ear. Handheld smartphone video, vertical 9:16, soft afternoon light.
Sound: light wind and distant seagulls. No music. No dialogue.

If you do not have good reference images yet, make them first. ChatGPT AI influencer photos shows how to build a reference set with one face.

What MiniMax H3 costs per clip

On MiniMax’s own API, H3 cost $0.08 per second at 768P and $0.13 per second at 2K on 5 October 2026. Reference images beyond the first five cost extra.

MiniMax-H3 output: “768P | Billed per second | $0.08 / second”, “2K | Billed per second | $0.13 / second”. MiniMax-H3-Max output: 480P $0.05 / second, 768P $0.08 / second. H3 input: “Audio | Free”, “Image | First 5 images free; $0.04 per additional image”, video billed by input duration at the output rate. H3-Max input: “Image | First 2 images free; $0.074 per additional image”. MiniMax-H3-Regeneration (768P to 2K): $0.05 / second.

MiniMax API docs — Pay as You Go pricing (opens in a new tab)checked 21 Oct 2026

Here is what one 10-second clip with six reference images costs, worked out from those prices:

Model on MiniMax’s API	Output, 10 s	6 reference images	Total
MiniMax-H3, 768P	$0.80	$0.04 (1 paid)	$0.84
MiniMax-H3, 2K	$1.30	$0.04 (1 paid)	$1.34
MiniMax-H3-Max, 768P	$0.80	$0.296 (4 paid)	about $1.10
MiniMax-H3-Max, 480P	$0.50	$0.296 (4 paid)	about $0.80

The “fast” Max model is not cheaper with a full reference set. It gives you only two free images, and each extra one costs almost twice as much. For persona work with five or more references, plain H3 at 768P is the better deal.

WaveSpeed sells two versions: MiniMax’s official API, and an “open-weights edition” that WaveSpeed runs on its own servers.

Open-weights edition per run: image-to-video $0.10 and reference-to-video $0.125, both shown as 50% off ($0.20 and $0.25). Image-to-video and reference-to-video with LoRA: $0.125 and $0.15. Official minimax/h3 image-to-video, text-to-video and reference-to-video: $0.70 per run each.

WaveSpeedAI — MiniMax H3 collection (opens in a new tab)checked 5 Oct 2026

The open-weights edition is the cheapest way to test. Two things to keep in mind. The price was a 50% discount on 5 October 2026, so it can double. And the open weights come with the licence covered further down. The WaveSpeed review explains how billing works there, and what generation costs lists H3 next to the other video models with dates.

MiniMax H3 makes its own sound

Every H3 clip has stereo sound: background noise, music and speech. You describe it in the prompt. None of the MiniMax pages read for this article shows a setting to switch it off.

The create-task reference on MiniMax’s API lists the model, the prompt, resolution, duration, ratio and the reference files. It has no parameter for sound on or off. So plan for it:

• For a silent clip or music, mute the H3 sound in your editor and add your own track.
• For a clip where she talks, the voice H3 makes up is new every time. Your fans will hear a different woman in every Reel. Give her one saved voice of her own and add it afterwards, or use a lip-sync model.
• Do not reuse someone else’s sound. On WaveSpeed, a reference video brings its soundtrack with it automatically. Only use clips you recorded yourself.

H3 can take an audio clip as a reference for a voice. The model card shows an example where a reference recording sets the “voice timbre” of a speaker. That is the tone of a voice, not a promise that she sounds exactly like your recording. Test it on your persona’s saved voice before you build on it.

Realistic speech also changes your labelling. On Instagram, a realistic voice needs a label on the post:

“We’ll require people to use this disclosure and label tool when they post organic content with a photorealistic video or realistic-sounding audio that was digitally created or altered, and we may apply penalties if they fail to do so.”

Meta Newsroom — Labeling AI-Generated Images on Facebook, Instagram and Threads (opens in a new tab)checked 5 Oct 2026

The AI label matrix shows what TikTok, YouTube and the others ask for.

MiniMax H3 open weights: the licence excludes the EU, UK, US and South Korea

The weights are free to download, but the public licence does not cover the European Union, the United Kingdom, South Korea or the United States. If you live there, you need a separate licence from MiniMax to run H3 yourself. This is not legal advice. It describes the published texts on 5 October 2026.

“MiniMax H3 release date/License date: August 2, 2026.” “‘Applicable Territory’ means worldwide, excluding the Excluded Territories.” “‘Excluded Territories’ means the European Union, the United Kingdom, the Republic of Korea and the United States of America.” Section II: “Solely within the Applicable Territory, we grant you a non-exclusive, non-transferable, royalty-free, limited license”. Section V.4: “You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.”

MiniMax H3 Community License Agreement (opens in a new tab)checked 5 Oct 2026

V.4 also names the “Outputs”. So the rule is about more than running the model. It also covers using or showing the clips outside the licensed area.

Where the licence does cover you, two more rules matter:

Section IV.1: “You shall obtain a separate, prior written authorization from MiniMax by contacting api@minimax.io with the subject line “MiniMax H3 licensing - authorization request”, if your commercial products and services generate more than 20 million US dollars (or equivalent in other currencies) in yearly revenue.” Section V.3: “You may not use the MiniMax H3 Works or any of their Outputs or results to improve any other artificial intelligence model (other than MiniMax H3 or its Model Derivatives).”

MiniMax H3 Community License Agreement (opens in a new tab)checked 5 Oct 2026

The revenue line is far away for one persona. V.3 matters sooner. If you train a LoRA?LoRA A LoRA is a small file trained on images of one subject. It teaches a general image model to reproduce that specific face or style. Open the glossary for another model, such as Wan, on clips made with H3, you use H3’s outputs to improve another model. So keep H3 clips out of training sets for other models.

MiniMax explains the split in a Q&A next to the weights:

Asked “Why can MiniMax-H3 API be used globally if open weights are restricted in some regions?”, MiniMax answers that for the API it controls the servers and can enforce safeguards, including “Protection against misuse involving minors”, “Copyright-related compliance measures” and “Content safety controls”. On restricted regions: “Yes. Organizations in these regions can apply for a formal license. After reviewing the deployment scenario and confirming that appropriate compliance controls and safeguards are implemented, MiniMax may authorize usage.”

MiniMax H3 — Q&A about the License (opens in a new tab)checked 5 Oct 2026

What that means depends on where you live and how you run H3:

Where do you live, and how will you run H3?

EU, UK, US or South Korea, via MiniMax's APISame countries, on my own GPUSame countries, a host's open-weights editionAnywhere else, on my own GPU

• Covered by the APIMiniMax says its API can be used globally. Its API terms apply. Label your clips as AI either way.
• Not covered by the public licenceApply to MiniMax for a licence before you run the weights.
• Not answered anywhereAsk the host in writing whether its licence covers your country. Until then, use the official API.
• The Community LicenseThe public licence applies. Among its rules: label machine-made content and do not impersonate anyone.

Schematicthe shape, not measured data

If you run your persona from Germany, the UK or the US, the short rule is: use MiniMax’s own API, and apply for a licence before you run the weights yourself.

Running MiniMax H3 in ComfyUI

ComfyUI?ComfyUI ComfyUI is a node-based interface for building image-generation pipelines. Each step is a box that is wired to the next. Open the glossary has native H3 workflows for text-to-video, image-to-video and reference-to-video. It needs several large files, and neither MiniMax nor ComfyUI publishes how much graphics memory you need.

Files: diffusion models minimax_h3_fl2va_pruned_int8_convrot.safetensors (text-to-video and image-to-video) and minimax_h3_ref2va_pruned_int8_convrot.safetensors (reference-to-video), the text encoder qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, a video VAE and an audio VAE, plus optional turbo LoRAs (8 steps for text- and image-to-video, 4 steps for reference-to-video). The text-to-video workflow is described as “Generate videos from text prompts with native stereo audio”.

ComfyUI Docs — MiniMax H3 native workflow (opens in a new tab)checked 5 Oct 2026

The model card says H3 has 33 billion parameters, and MiniMax’s own serving example uses four graphics cards:

“H3-Omni-Transformer is a 33B-parameter dense, single-stream Transformer”. SGLang deployment example for reference-to-video: sglang serve --model-path MiniMaxAI/MiniMax-H3 --num-gpus 4 --ulysses-degree 4 --performance-mode speed --host 0.0.0.0 --port 30011 --model-variant ref2va.

Hugging Face — MiniMax-H3 model card (opens in a new tab)checked 5 Oct 2026

ComfyUI’s int8 and nvfp4 files are compressed versions that need less memory, but no page names a minimum card. Check the file sizes against your card before you download, or rent a large GPU. In the EU, UK, US or South Korea, a local install is also not covered by the public licence, and the ComfyUI page does not mention that. Test through the API first: a 5-second clip at 768P costs $0.40 there.

What MiniMax’s terms ask of you: rights and labels

MiniMax’s API terms make you promise that you own what you upload, and they let you keep the rights to your clips. Both the terms and the licence want AI content marked.

Effective Date: March 30, 2026. “You guarantee that any content (referred to as “Client Content”) you provide, publish, upload, or process while using the products and services […] are owned by you or that you have obtained legal authorization from the rights holders.” “As between you and us, and to the extent permitted by applicable laws, you retain your ownership rights in Client input and generated content. We may use the input and generated content to provide, maintain, develop, and improve our Services”. “The services provided by us are based on deep synthesis technology. If you offer services based on these technologies, you must […] place a prominent mark in reasonable positions or areas within generated or edited content to inform the public of the use of deep synthesis technology.”

MiniMax — Open Platform Terms of Service (opens in a new tab)checked 5 Oct 2026

The licence lists among the forbidden uses: posting machine-generated content in public “without clearly and prominently disclosing that such information and/or content is machine-generated”, and using H3 “to impersonate another person without that person’s consent”.

So upload only your own generated persona, and label every realistic clip as AI. Keep the files as they come out: don’t strip your watermarks shows how to check what a clip carries before you post. The risk guide covers what you can protect once the clip exists, and what to do when someone copies it.

Start with one test. Take your five best reference images of her, write the harbour prompt above, and run it three times as a 5-second clip at 768P on MiniMax’s API. That costs $1.20. If her face holds to the last frame in two of three runs, H3 is worth a place in your video workflow. Then compare it with the other models in the AI video model comparison. If you want her to copy a movement you filmed, Kling motion control does that job better than a written prompt.