How to prompt MiniMax H3 in OpenModels

Start with the shot you want a viewer to see, then explain what changes and what they hear. This guide turns MiniMax's published formats into a practical drafting workflow, with original examples and a clear distinction between subject references and frame controls.

In short

Select the H3 model and attach the references you intend to use before enhancing your prompt.
H3 Studio uses separate description, soundscape and music fields; the hosted H3 lane produces silent video.
OpenModels H3 reference inputs guide the subject and motion feel of a new scene; they are not first or last frames.
Read the enhanced draft, check the quote and press Generate only when you are ready.

Who can run this: Reading is free. Signed-in H3 Studio generation can use GPU credits; hosted H3 uses Premium model credits. The composer shows your available options and quote.

Which plan: The guide is available without an account. To render, sign in and check the selected model's access, length, resolution and quote. H3 Studio can use GPU credits; the hosted H3 option uses Premium model credits. Free-plan video has a resolution limit.

The two official guides and what they cover

MiniMax's base prompting guide describes a shot-oriented visual description with separate ambient sound and score fields. It uses numbered shots, explicit camera movement and consistent speaker labels. Dialogue keeps its original words. Its upstream examples also cover image and endpoint conditioning; those examples describe model modes, not a promise that every application exposes those inputs.

The full-reference guide adds a way to name source material, explain what should carry over, and describe the resulting sequence. Its six section labels are subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape and non_diegetic_music. Use the original documents for the complete notation. The examples below are our own, written for the controls OpenModels currently exposes.

Choose the workflow before writing the prompt

Your selected workflow	What to write	What to check
H3 Studio, text only	A visual sequence, physical sounds and an optional score.	Keep the scene short enough for the chosen duration.
H3 Studio, references attached	A new scene featuring the referenced people, objects or motion feel.	Assign each attachment a clear role; references do not fix the opening frame.
Hosted H3	A visual description that stands on its own.	This OpenModels lane produces silent video.
Exact first or last frame needed	A transition between the frame inputs of a compatible model.	Choose a model with the relevant frame slots, such as the LTX first/last-frame workflow.

Start by deciding whether you need a familiar subject, an exact opening composition or a continuation of existing footage. Those are different jobs. A portrait reference can tell a model whom to depict without placing that portrait at the beginning of the result. Likewise, a motion reference can inform movement without turning the task into video editing. Making that decision before drafting saves you from polishing instructions aimed at the wrong input type.

Watch out: In OpenModels, H3's References workflow creates a new shot. Do not ask it to continue, extend or edit the attached video, or treat an attached picture as a guaranteed first or last frame. Select a dedicated workflow when that is the result you need.

Write a first draft in five passes

Name the deliverable Write one sentence explaining the result: a teapot product shot, a character reaction or an establishing view. Choose one thing that must be readable in the final clip. That gives you a useful test when reviewing the output instead of relying on a vague judgement that it should feel more cinematic.
Stage the opening State the subject's location and the framing. A cup already on the table is simpler to reason about than a cup which must be introduced later. Identify which objects are essential, and leave incidental decoration out until the scene works. Check that attached reference files are the files you actually want to use.
Describe the visible change Write the action in order: lift, pour, replace. Give the subject a clear finish so the clip has a destination. Avoid turning a short product moment into a list of travel, wardrobe changes, camera cuts and dialogue. If a beat could be its own clip, consider drafting it separately.
Choose the camera and audio Pick a camera move that helps the viewer see the action, then decide which sounds matter. A close view of liquid pouring needs different framing from a wide view of someone setting a table. For Studio, specify whether the scene needs a score. For hosted H3, focus the draft on visual information.
Enhance and review Select the intended model and add the real references before using Enhance prompt. Read the rewritten draft for changes to the subject, action and attachment roles. Keep a copy of your short brief so you can judge whether the expansion preserved your idea. Check the generation quote before submitting.

An original H3 Studio example

Here is a small product scene with one continuous action. It has an observable beginning, middle and end: the pot rests on the table, the drink is poured, and the pot returns. There is no required dialogue or music. Choose a duration that leaves time for the action rather than assuming a prompt can override the length selected in the composer.

integrated_multimodal_description: [Shot 1] Live-action tabletop product film. A small orange enamel teapot sits beside an empty white cup on a pale wooden table. A hand enters from the right, lifts the teapot and pours a thin stream into the cup. The camera pushes in with small amplitude at slow speed. Warm window light catches the rising steam. The hand returns the teapot to its original place.

overall_soundscape: A gentle pour is followed by one soft ceramic tap when the teapot returns to the table. Quiet room ambience continues underneath.

non_diegetic_music: N/A

The teapot example is deliberately easy to revise. If the pour is the important moment, bring the framing closer and leave the rest alone. If you need a wider advertising composition, keep the objects and action but ask for empty space on one side. If a brand mark matters, prepare a suitable reference and review the result carefully; a detailed prompt is guidance, not a guarantee that every small visual detail will survive.

Turn a reference collection into a clear brief

Before attaching several files, write an ordinary-language inventory. For example: image one supplies the teapot design, image two supplies the table setting, and the video supplies the pace of the pouring gesture. That inventory makes conflicts visible. If one image shows a green pot and another shows an orange pot, decide which design wins instead of asking the model to resolve the contradiction for you.

A useful reference draft for this example would be: Create a new tabletop scene with the teapot from the first attached image and a table arrangement inspired by the second. Keep the teapot's shape and orange enamel. A hand pours into one white cup, pauses, then replaces the pot. Use the attached clip only as a guide to the gentle pace of the gesture. Frame the scene from a low three-quarter angle with warm window light.

That is a drafting example, not a promise to reproduce a source frame. After enhancement, inspect the attachment references and the descriptions of what should carry over. A longer rewrite is useful only if it makes the instructions easier to follow. Remove invented requirements that distract from your deliverable, particularly an extra performer, location change or camera move you did not ask for.

Review the result with one question at a time

Subject: Is the important object or character identifiable throughout the action? If not, simplify the scene and make its reference role clearer before adding more decoration.
Action: Can you tell what happens and in what order? If the clip feels hurried, remove a beat or select an appropriate available duration instead of repeating the same instruction in stronger language.
Framing: Does the camera reveal the part of the action you care about? A beautiful wide view can still be the wrong result when the task requires a close view of hands.
Audio: On Studio, listen for the intended physical event and whether the score obscures it. On the silent hosted lane, plan any soundtrack as a separate editing step.
Revision: Change the instruction most directly connected to the defect. Keep the working subject, composition and reference choices so the next comparison is useful.

Keep a small record of the draft, model, attachments and chosen settings next to each result. You can then distinguish a wording revision from a change of workflow or duration. This is especially helpful when comparing two versions of a scene: the useful question is which instruction helped the deliverable, not which draft contained the most adjectives. Every new render is a separate generation, so review the displayed quote when deciding whether another variation is worthwhile.

Questions about H3 prompts

Do I need to write the native sections by hand?

You can begin with a plain-language brief. Select the model and attach the intended inputs before using Enhance prompt, then inspect the structured draft. Knowing what each part is for helps you review it; adding more labels without a clear scene does not fix an unclear brief.

Can I use an H3 reference as the exact first frame?

The OpenModels H3 References workflow uses it to guide a new scene, not to anchor an exact first frame. For a fixed opening or ending, choose a model and workflow that expose those frame controls. The upstream H3 documents describe additional modes beyond this reference workflow.

Why is there no sound in my hosted H3 result?

The hosted H3 lane in OpenModels produces silent video. H3 Studio supports video with sound. Select the workflow that matches your deliverable before writing audio instructions, and check the model's current options in the composer.

Can I use the H3 format with other models?

Keep the scene idea, but adapt it to the selected model. Other video workflows may expect a continuous visual description, a supplied starting frame or two endpoint images. Use the model's own short guide and Enhance prompt after selecting it instead of carrying H3 section labels into every workflow.

Open the video composer — Opens the studio. Nothing runs, and nothing is charged, until you press Generate.

Last checked 2026-09-20. Prices and plan limits change; the pricing page is the authority.