> For the complete documentation index, see [llms.txt](https://docs.dreamerland.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.dreamerland.ai/video/minimax-h3.md).

# Minimax H3

MiniMax H3 is MiniMax's latest omni-modal video model, and it's now fully available in Dreamerland. Unlike earlier video models that only read a text prompt or a single image, H3 understands **text, images, video and audio together** — you can hand it a prompt, a few reference photos, a video clip and a music track all at once, and it fuses them into a single coherent shot.

Every H3 clip comes with **native audio generated together with the picture** — dialogue, sound effects and ambience are created in the same pass, so there's no separate audio step and no audio toggle to worry about. Clips run from **4 to 15 seconds**, and you can chain them together with the Extend task to build longer scenes.

### Choosing a Task

There are 6 Minimax tasks in Dreamerland:

* Text to Video
* Image to Video
* [Reference to Video](/video/minimax-h3/reference-to-video.md)
* Extend Video
* [Video Inpaint](/video/minimax-h3/video-inpaint.md)
* [Music Video](/video/minimax-h3/music-video.md)

{% tabs %}
{% tab title="Text" %}
**Text to Video**

The classic starting point — describe the scene and H3 generates it from scratch, sound included. Because H3 generates audio natively, it pays off to describe what you want to *hear* as well as see.
{% endtab %}

{% tab title="Image" %}
**Image to Video**

Bring a still image to life. Upload a **start frame**, an **end frame**, or both:

* **Start frame only** — the video begins from your image and animates forward.
* **End frame only** — H3 generates a shot that lands exactly on your image.
* **Both** — H3 creates a smooth transition between the two frames. Great for morphs, reveals and before/after shots.
  {% endtab %}

{% tab title="Reference" %}
[**Reference to Video**](/video/minimax-h3/reference-to-video.md)

The most powerful H3 task. Instead of using an image as the literal first frame, you feed H3 **references** and it builds a new shot around them:

* **Up to 5 reference images** — lock a character's face, an outfit, a product or an art style.
* **A reference video** — borrow motion, choreography or camera work from an existing clip.
* **A reference audio** — drive the clip with a voice recording or sound. On H3, an audio reference alone is a valid job: upload a voice line and let the model perform it.
  {% endtab %}

{% tab title="Extend" %}
**Extend Video**

Continue a video you've already made. Upload the previous clip, choose how the extension should be driven (**text**, **first/last frame**, or **reference**), and H3 picks up where the clip left off — the extension automatically inherits the source clip's aspect ratio.

Turn on **Merge source video** if you want the result delivered as one combined clip (original + extension); leave it off to receive just the new part.

This is how you build stories longer than 15 seconds: generate a shot, extend it, extend again.
{% endtab %}

{% tab title="Inpaint" %}
[Video Inpaint](/video/minimax-h3/video-inpaint.md)

Edit an existing video with a text instruction — no manual masking needed. Upload the source clip and type a **mask prompt** describing what to change, e.g. *"replace the red car with a white horse"* or *"change her jacket to a leather one"*.

* **Control type — None:** the region is freely regenerated from your prompt.
* **Control type — Pose:** the original motion is preserved while the appearance changes — perfect for swapping characters or outfits without losing the performance.

You can also attach up to 5 reference images to tell H3 exactly what the replacement should look like. The output keeps the source clip's length and aspect ratio.
{% endtab %}

{% tab title="Music" %}
[**Music Video**](/video/minimax-h3/music-video.md)

Purpose-built for making videos cut to a song. Upload a full track (up to 5 minutes) and H3 generates a clip synchronized to it — visuals that move with the music.

Because each generation is still one 4–15 second clip, you build the full video piece by piece:

1. Upload your song and generate the first clip (elapsed length = 0).
2. For the next clip, set **Elapsed video length** to how many seconds you've already covered (e.g. 10).
3. Optionally attach the **previous clip** so the new one continues seamlessly from it — same characters, same scene, next part of the song.
4. Repeat until the song is covered, then stitch the clips in the built-in video editor.

Reference images and a reference video are optional here too, so your performer can stay consistent across every cut.
{% endtab %}
{% endtabs %}

### Official Prompt Guide

You can find the official prompt guide below:

Text to Video and Image to Video: <https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md>

Reference: <https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/docs/VIDEO_PROMPT_WRITING_GUIDE_ref_en.md>

You can also use our [AI Director](/video/minimax-h3/ai-director-kumi.md) to help you create prompt following official prompt guide