MiniMax H3 Prompt Guide: Better Video, Audio & Camera Results

Core MiniMax H3 Prompt Structure and Syntax

MiniMax H3 prompt structure differs by workflow. Base text and keyframe modes use 3 named sections; full-reference Ref2VA uses 6 named sections.

Base Mode Sections (T2VA, I2VA, FL2VA, L2VA)

MiniMax H3's base prompt sections are defined in the official SKILL.md documentation:

• integrated_multimodal_description — the complete visual and motion description: composition, subjects, environment, actions, and camera behavior, written as a continuous timeline.
• overall_soundscape — all audible environmental elements, including ambient noise, room tone, and diegetic sound.
• non_diegetic_music — score mood, instrumentation, tempo, and the point at which music enters relative to on-screen action.

Ref2VA Sections

MiniMax H3's Ref2VA prompt uses 6 sections in this order:

• subject_definitions — names and describes each reference input using consistent labels, such as <Picture 1>, <Video 1>, and <Audio 1>.
• summary — a brief statement of the overall scene and creative intent.
• retention_analysis — specifies what visual or audio qualities should be preserved from each reference.
• detailed_description — the full shot-by-shot timeline, including where each reference appears.
• overall_soundscape — environmental audio.
• non_diegetic_music — score instructions.

The following annotated example shows a complete MiniMax H3 T2VA prompt:

integrated_multimodal_description:

A woman in a white lab coat examines a glowing petri dish in a dim laboratory with blue neon accents. The camera makes a slow close-up push toward her face as she lifts the dish to the light. The action unfolds continuously over five seconds.

overall_soundscape:

A quiet electrical hum, soft glass handling, and distant laboratory ventilation.

non_diegetic_music:

One low, restrained piano note begins as the camera reaches the close-up.

MiniMax H3's official skill recommends concrete visual and audio detail over generic praise words such as "cinematic" or "beautiful." The described timeline should match the requested clip duration.

Camera and Shot Control Language for MiniMax H3

MiniMax H3 controls camera output when you state a single coherent plan — shot size, movement, and framing — as a dedicated clause inside integrated_multimodal_description. Stacking contradictory instructions in one short clip produces unstable motion.

MiniMax H3's video guide supports direct motion instructions such as [pan], [zoom], and [static] placed after key descriptions. Practical camera language can also include shot size, direction, and framing.

Shot Size

• Wide shot — establishes environment; subject occupies less than one-third of frame height.
• Medium shot — frames subject from waist up; standard for dialogue and action sequences.
• Close-up — isolates face or object detail; fills the frame with a single focal element.
• Extreme close-up — restricts frame to one feature, such as an eye or a hand.

Camera Movement

• Pan — horizontal rotation from a fixed position; write "slow pan left" or "fast pan right."
• Tilt — vertical rotation from a fixed position; write "tilt up to reveal the skyline."
• Dolly / forward move — physical forward or backward camera travel; write "slow forward camera move."
• Tracking shot — lateral travel that follows a moving subject; write "tracking shot following the runner."

A clear camera clause combines these in sequence without contradiction. For example: "Medium shot; slow forward camera move; centered composition." This gives MiniMax H3 three resolved parameters with no ambiguity.

Using Reference Images and First-Frame / Last-Frame Prompting

MiniMax H3 anchors a generation to one or two supplied images in I2VA and FL2VA modes. There are 3 distinct reference-image workflows MiniMax H3 supports:

1. I2VA (Single First-Frame Image)

MiniMax H3's I2VA workflow takes one image as the first frame and generates all subsequent frames from it. The prompt describes motion and environment changes, not the subject's appearance. Write the motion clause inside integrated_multimodal_description, describing what changes from the reference frame forward.

2. FL2VA (First and Last Frame)

MiniMax H3's FL2VA workflow accepts both a first-frame image and a last-frame image, then synthesizes the transition between them. The prompt's job is to narrate the motion arc connecting the two images: what moves, in what direction, and at what pace.

MiniMax H3 worked example — FL2VA transition prompt:

integrated_multimodal_description:

<Picture 1> shows a woman standing by a bright window; <Picture 2> shows the same woman seated at a desk after dusk. The camera holds a wide composition as she walks from the window to the desk and sits. Daylight fades steadily into cool evening light without a cut.

overall_soundscape:

Soft morning room tone transitions gradually to an evening hum.

non_diegetic_music:

None.

3. Ref2VA (Reference Images, Video, and Audio)

MiniMax H3's Ref2VA workflow accepts reference images, videos, and audio in a single request. Keep reference labels identical in every section — for example, <Picture 1> in subject_definitions and again in detailed_description — and state exactly what should be retained from each reference.

Important: First/last-frame and reference workflows are mutually exclusive. Do not mix a first_frame or last_frame item with reference_image, reference_video, or reference_audio in the same MiniMax H3 request.

MiniMax H3's subject consistency across all workflows depends on keeping the reference image's subject unobstructed and well-lit. A cropped or partially occluded reference produces drift in generated frames.

Voice, Dialogue, Sound, and Music Prompting in MiniMax H3

MiniMax H3 generates audio alongside video. You specify ambient sound and music using the two dedicated audio sections: overall_soundscape and non_diegetic_music. No separate audio track or configuration file is required.

There are 4 distinct audio element types you write into a MiniMax H3 prompt:

• Ambient sound cues — environmental audio written in overall_soundscape, such as "rain drumming on a tin roof" or "crowded café background noise."
• Diegetic sound — on-screen sound events tied to action, described inside integrated_multimodal_description at the moment they occur.
• Music cues — genre, tempo, or instrument instructions in non_diegetic_music, such as "sparse jazz piano fades in at the end of the line."
• Reference audio — in Ref2VA, label the audio reference consistently, such as <Audio 1>, and describe how it should influence the result in retention_analysis and detailed_description.

MiniMax H3 syncs audio to on-screen action by referencing the visual event as the trigger inside the prompt. Write "a low cello note rises only as the detective turns" rather than placing the music cue in isolation. The model uses the visual event description as the synchronization anchor.

MiniMax H3 audio example — cinematic scene:

overall_soundscape:

Steady rain, distant traffic, and a faint electrical buzz from the sign.

non_diegetic_music:

A low cello note rises only as the detective turns.

Copy-Ready MiniMax H3 Example Prompts by Use Case

MiniMax H3 example prompts below are copy-paste starters you can adapt immediately for cinematic film scenes, talking-head videos, and product showcase shots. Each follows the three-section base structure from the official SKILL.md.

Cinematic Film Scene

integrated_multimodal_description:

A detective stands at the entrance to a rain-soaked alley before dawn. Medium shot; slow forward camera move. As the detective turns toward a flickering sign, rain gathers on the coat collar and reflected neon trembles in the puddles.

overall_soundscape:

Steady rain, distant traffic, and a faint electrical buzz from the sign.

non_diegetic_music:

A low cello note rises only as the detective turns.

Swap: Replace "detective" and "alley" with your subject and location. Replace "cello note" and "sign buzz" with your target audio cues. The visual event ("detective turns") anchors the music trigger.

Influencer / Social Talking-Head

MiniMax H3 handles talking-head formats well when the prompt specifies tone and framing clearly.

integrated_multimodal_description:

A confident woman sits at a minimalist white desk facing camera, natural window light from the left. Static medium close-up; slight rack focus from background to face at the start. She speaks directly to the lens in a warm, conversational tone throughout the clip.

overall_soundscape:

Faint café noise in the background. No additional ambient elements.

non_diegetic_music:

None.

Swap: Replace "woman" with your talent description. Replace "white desk" with your set dressing. Add spoken dialogue inside integrated_multimodal_description using a speaker tag and quoted text.

Product Showcase

integrated_multimodal_description:

A matte-black wireless headphone rests on dark concrete. The camera makes a slow orbit while a rim light reveals its texture. The product remains centered and the motion completes in five seconds.

overall_soundscape:

A clean studio-room tone with a quiet mechanical turntable sound.

non_diegetic_music:

One short electronic chime at the end of the orbit.

Swap: Replace "matte-black wireless headphone" and "concrete surface" with your product and surface material. Replace "electronic chime" with a brand-appropriate audio cue.

Each prompt follows the integrated_multimodal_description → overall_soundscape → non_diegetic_music structure from MiniMax H3's official SKILL.md guidance. Adapt templates by changing one variable at a time — subject first, then environment, then camera move.

MiniMax H3 Quick-Reference Cheat Sheet: Prompt Template Table

This table maps every prompt element to what it controls, a ready-to-copy phrasing pattern, and the workflows where it applies.

This MiniMax H3 cheat sheet works best as a build-order guide. Fill each row with your own content, then stack the sections in the order the official skill specifies for your chosen workflow.

Common MiniMax H3 Prompting Mistakes and How to Fix Them

The most frequent MiniMax H3 prompting mistake is writing a vague, caption-like prompt that describes what a scene looks like rather than what happens during the 4–15-second clip. The fix is to lead with dynamic action and state what changes across the clip's full duration.

MiniMax H3 prompt failures cluster into 5 recurring patterns:

1. Vague Subject Description

Writing "a woman walking" gives the model no physical anchor.

Fix: Add appearance, clothing, and environment inside integrated_multimodal_description.

2. Contradictory Camera Cues

Stacking "wide shot, extreme close-up, pan left, and static camera" in one short clip forces the model to arbitrate between incompatible instructions.

Fix: Commit to exactly one primary shot type and one camera movement per prompt.

3. Missing Audio Direction

Omitting the overall_soundscape and non_diegetic_music sections does not produce silence — it produces unpredictable ambient noise.

Fix: Complete both audio sections explicitly, even if the entry is "None."

4. Ignoring Frame Anchoring in I2VA Mode

Submitting a reference image without a prompt that describes the motion extending from that frame causes the model to treat the image as a static poster.

Fix: Describe the first visible action that departs from the reference frame.

5. Mixing First/Last-Frame and Reference Inputs

These workflows are mutually exclusive in the MiniMax H3 API.

Fix: Choose one workflow per request and structure the prompt accordingly.

Where and How to Access MiniMax H3

MiniMax H3 is accessible through 2 primary routes: the MiniMax video API and multi-model platforms that have integrated the model.

MiniMax H3's video API is available at POST https://api.minimax.io/v2/video_generation using the model name MiniMax-H3. A valid request requires a non-empty text item in content, a resolution, and a duration of 4–15 seconds. Text-to-video requests also require an explicit aspect ratio.

Multi-model platforms aggregate MiniMax H3 alongside other engines, letting creators switch models without leaving a single interface. Minimax H3 is one platform for accessing MiniMax H3 alongside other video and image engines in one workspace.

Pricing for MiniMax H3 varies by tier and platform. A dedicated pricing breakdown covers credit costs, free generation limits, and subscription options in full — refer to that article for current figures before committing to a workflow.

Where to Go from Here

MiniMax H3 prompt writing rewards correct workflow selection and consistent section structure. Choose the right workflow first — T2VA, I2VA, FL2VA, L2VA, or Ref2VA — then fill the appropriate named sections with concrete visual and audio detail.

The next practical step is to run the copy-ready templates against a real project. Iterate on overall_soundscape and non_diegetic_music first. These 2 sections produce the most visible gains per revision because they are the most commonly omitted.

Creators who want a single workspace to test MiniMax H3 prompts alongside other leading models may find a multi-model platform useful. Explore Minimax H3 to test the workflow alongside other video and image engines.

Frequently Asked Questions

How do you write a good prompt for MiniMax H3?

A good MiniMax H3 prompt starts by choosing the correct workflow. Then use the appropriate named sections — integrated_multimodal_description, overall_soundscape, and non_diegetic_music for base modes; six sections for Ref2VA. Fill each section with concrete details about composition, subjects, environment, action, camera, and sound.

What is the difference between T2VA and I2VA in MiniMax H3?

MiniMax H3 supports two base text workflows: T2VA creates a full audiovisual timeline from text alone. I2VA begins with an uploaded first-frame image and develops forward from it, so the prompt focuses on motion and camera direction rather than scene construction.

Can MiniMax H3 use first and last frames together?

Yes. MiniMax H3's FL2VA workflow accepts a text prompt with both a first-frame image and a last-frame image. The prompt describes the continuous motion arc connecting the two anchors.

Can MiniMax H3 use audio references?

Yes. MiniMax H3's Ref2VA workflow accepts reference audio alongside reference images and video. Label each audio reference consistently — for example, <Audio 1> — and describe how it should influence the result.

How do I control the camera angle and shot type in a MiniMax H3 prompt?

Camera control in MiniMax H3 is achieved by placing a coherent camera description inside integrated_multimodal_description. MiniMax's guide supports direct motion instructions such as [pan], [zoom], and [static], as well as natural-language shot descriptions. State one coherent plan — shot size, movement, and framing — without contradictory instructions.

Can I mix first/last-frame inputs with reference images in one request?

No. First/last-frame and reference workflows are mutually exclusive within one MiniMax H3 API request. Do not combine first_frame or last_frame items with reference_image, reference_video, or reference_audio in the same request.

Where is the official MiniMax H3 prompt documentation?

The official prompt structure documentation for MiniMax H3 is published in the SKILL.md file inside the MiniMax-AI H3 repository.