MiniMax H3 Prompt Guide for Video, Camera, and Sound

A strong MiniMax H3 prompt is less like a bag of visual adjectives and more like a compact directing brief. H3 can combine text with image, video, and audio references, so the prompt must explain both what to create and how each reference should influence the result. The goal is not maximum length. The goal is clear relationships, timed action, coherent camera language, and an intentional soundscape.

This guide provides a reusable structure for text-to-video, first-and-last-frame, and multimodal reference generation. It reflects the input roles and limits documented by the official MiniMax H3 V2 API.

The seven-part H3 prompt structure

Use this order as a starting point:

1. Format and purpose: product commercial, documentary insert, music visual, cinematic scene, or social clip.
2. Subject and setting: who or what is present, where, time of day, and important visual attributes.
3. Action beats: what happens first, next, and last.
4. Camera: shot size, lens feeling, angle, movement, focus, and cuts.
5. Look: lighting, palette, texture, production design, and realism level.
6. Dialogue and sound: spoken words, voice quality, effects, ambience, music, and timing.
7. Constraints: what must remain consistent or must not change.

Write in complete, concrete sentences. If several instructions conflict, the model must decide which one wins.

A text-to-video example

Weak prompt:

Directed prompt:

The second prompt gives H3 a sequence, not just a mood.

Control time with beats

For clips up to 15 seconds, use approximate time anchors or an ordered sequence. Do not request twelve unrelated events in four seconds. A useful pattern is:

• 0–3 seconds: establish the subject.
• 3–8 seconds: perform the main action.
• 8–12 seconds: reveal or transformation.
• 12–15 seconds: resolve on a clear final image.

Multi-shot prompts should state where a cut occurs. If you want one continuous take, say so and avoid verbs such as “cut to.” Continuity becomes easier when each beat grows naturally from the previous one.

Write usable camera direction

Combine one primary movement with a clear subject relationship: “slow dolly toward the cyclist while the camera stays at wheel height.” Terms such as tracking, orbit, crane, handheld, static, rack focus, and push-in are useful when the surrounding sentence makes their purpose clear.

Avoid stacking incompatible directions: “static handheld drone shot with a rapid slow push-in.” Decide whether the camera or subject creates the energy. Specify lens character only when it changes the shot: macro compression, wide-angle spatial exaggeration, shallow portrait focus, or long-lens background compression.

Prompt dialogue and audio together

H3 can generate synchronized stereo audio, so write sound as part of the scene. Put exact dialogue in quotation marks and identify the speaker. State delivery, language, and timing:

Separate dialogue, sound effects, ambience, and music. If the scene should remain silent, say which sounds are allowed rather than only writing “no audio.” For reference audio, explain whether H3 should preserve voice identity, rhythm, ambience, or musical character.

First-frame prompting

The first frame already establishes composition and appearance. Do not spend the prompt redescribing every visible pixel. Focus on motion and preservation:

The official API determines image-to-video aspect ratio from the input frame, so use an image already composed for the intended output.

First-and-last-frame prompting

Treat the prompt as the motion bridge between two fixed states. Describe the causal transition instead of repeating the images:

Choose endpoint images with compatible subject design and perspective. Prompting cannot always reconcile fundamentally different geometry in a short duration.

Multimodal reference prompting

The V2 API uses explicit roles for reference images, videos, and audio. In the prompt, name the job of each asset:

This prevents accidental transfer. If a reference contributes only motion, say what should not be copied.

Negative instructions and constraints

Constraints work best when they are specific and limited. Useful examples include:

• Keep the product shape and logo placement unchanged.
• One continuous shot; no cuts or time jumps.
• No additional people enter the frame.
• Preserve the original dialogue; replace only ambience.
• Do not copy the background from the movement reference.

A long list of generic negatives can dilute the main direction. Prioritize the two or three failures that would make the result unusable.

Why prompts fail

Common causes include too many events for the duration, contradictory camera terms, unclear reference roles, a subject that changes description midway, dialogue without a named speaker, and sound instructions buried inside visual adjectives. Another frequent problem is demanding exact text, anatomy, complex interaction, and rapid motion simultaneously. Simplify the first attempt and add complexity after the core shot works.

Use a fixed seed and modify one section of the prompt at a time. If motion is wrong, revise beats and camera—not color adjectives. If identity drifts, strengthen preservation language or improve the reference image.

A copyable template

[Duration and format]. [Subject] in [setting].

Action beats: First, [...]. Then, [...]. Finally, [...].

Camera: [shot size, angle, movement, focus, cuts or continuous take].

Look: [lighting, palette, material, realism/style].

References: Use [asset] for [specific role]. Preserve [...]. Do not transfer [...].

Sound: [dialogue], [effects], [ambience], [music], [timing and stereo placement].

Constraints: [two or three critical requirements].

Test it in the playground

Every mapped example above includes Copy Prompt and Try This Prompt. Copy keeps the full directing brief on your clipboard. Try This Prompt opens the MiniMax H3 playground, inserts the exact mapped prompt, and scrolls to the input area. The videos are inspiration mappings rather than deterministic reproduction promises: generation can vary, and source examples may have used references or settings beyond the visible text. Start from one prompt, then change a single layer—subject, action, camera, look, or sound. For interface-specific help, read Prompting in the MiniMax H3 workspace.

Primary sources and status

• MiniMax H3 V2 API and input roles
• MiniMax H3 model card

The documented input roles and limits were last verified on August 7, 2026. The seven-part structure and example prompts are independent editorial guidance rather than an official MiniMax prompt formula.

There is no universally perfect prompt. The best prompt is the shortest directing brief that makes the intended sequence and reference relationships unambiguous.

Continue with the right workflow

• Build a clean first request with the Text to Video guide.
• Control endpoints with First and Last Frame.
• Assign image, video, and audio roles with Multimodal Reference.
• Plan motion transfer with the Reference-to-Video guide.
• Protect recurring subjects with the Character Consistency guide.
• Turn the structure into an advertising brief with Product Video Prompts.
• Write a natural vertical testimonial with UGC Video Prompts.
• Build a multi-shot sequence with the Music Video Workflow or Game Cinematic Prompts.
• Paste any example prompt into the MiniMax H3 Playground and change one directing layer at a time.