Published ratesView pricing →

MiniMax H3 · Prompting

MiniMax H3 prompt guide: schema, notation and a prompt library

Most MiniMax H3 prompts come out flat because they are written as a description when the model was trained on a schema — three named fields, timed shots, tagged dialogue. This MiniMax H3 prompt guide covers that schema, the notation that goes inside it, and 24 complete prompts written in it that you can copy or preview.

Most people arriving at a MiniMax H3 prompt guide want prompts, not theory — the library is one jump below. The sections before it explain why every MiniMax H3 prompt here is shaped the way it is.

Anatomy of an H3 promptT2VA · one shot
00:00End of clip
Picture[Shot 1] Live-action, cinematic, a medium-wide shot frames a silver-haired violinist in a rain-soaked wool coat standing under a shop awning at night. The camera pushes in with small amplitude at slow speed as she finishes a phrase, lowers the bow, and turns toward the doorway behind her, ending on her face in close-up.
SoundRain drums steadily on the canvas awning above a low street ambience. Water runs from the edge of the fabric onto the pavement, and a door hinge creaks once behind her.
ScoreA solo violin line at a slow tempo, warm and slightly detuned, thinning to a single sustained note before it fades.

All three come out of one forward pass. Leave the bottom two empty and the model still fills them — silence is a thing you ask for, not a thing you get by saying nothing.

Prompts
24
Categories
8
Fields
3
Modes
4

MiniMax's own spec

What MiniMax's own prompt guide specifies

MiniMax ships a prompt-writing document with the model, and it asks for something more structured than a paragraph: An image-alignment instruction on the first line, one blank line, then the three core fields. Most of the MiniMax H3 prompt guide articles on the first page of results were written from each other; this one starts from that file.

Pictureintegrated_multimodal_description
The visual timeline: subject, what changes over the clip, and the camera describing it. MiniMax asks for camera motion written as a natural English action inside the shot, not stacked as labels.
Soundoverall_soundscape
One continuous paragraph, 1–4 sentences, covering ambient sound, physical action sounds and non-verbal human sounds. Dialogue and singing are explicitly excluded from this field.
Scorenon_diegetic_music
1–3 sentences for score the characters cannot hear. MiniMax asks for instrumentation, speed, rhythm and dynamic changes — and explicitly tells you not to use abstract mood words.

The split between the last two fields is the part worth internalising: sound the characters can hear and score they cannot are separate fields, and MiniMax asks you to describe the score by instrumentation and rhythm rather than by mood.

Source: MiniMaxAI/MiniMax-H3 · docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md

MiniMax's own spec

The four modes, and the first line each one needs

MiniMax labels the modes T2VA, I2VA, FL2VA and L2VA — text, first-frame, first-and-last-frame, and last-frame, each producing video and audio together. ComfyUI's template names (Text to Video, Image to Video, Reference to Video) map onto the same models. Every MiniMax H3 prompt except the text-only one opens with a fixed alignment instruction that tells the model which frame a reference image belongs to. These are worth copying verbatim rather than paraphrasing — a MiniMax H3 prompt guide that rewrites them in its own words is giving you a paraphrase of a contract.

  • T2VA

    Text only. Builds the whole audiovisual timeline from the prompt.

    First line

    No alignment instruction — start directly with integrated_multimodal_description.

  • I2VA

    One image as the opening frame, then development forward from it.

    First line

  • FL2VA

    First and last frame. The body supplies the motion path between them.

    First line

  • L2VA

    Last frame only. Infer a plausible earlier state and converge onto the image.

    First line

S.SS is the effective clip duration to exactly two decimal places, and N is the index of the final shot. The instruction is the first line, followed by one blank line before the fields.

MiniMax's own spec

The notation MiniMax H3 prompts actually use

This is the part almost every MiniMax H3 prompt guide skips. The official document does not just give advice — it defines a small notation for shots, speakers, dialogue and on-screen text. Prompts written without it work; prompts written with it repeat.

[Shot 1]
Opens a shot. The first shot carries no timestamp; every later shot starts with a strictly increasing cut time inside the clip duration.
[Shot 2] At 00:03.500, the camera cuts to a close-up of the kettle.
(S1) (S2)
Stable speaker IDs. A speaker keeps the same ID across shots; characters who never vocalise get no ID. Use (S1,S2) when they speak together.
The middle-aged baker with a calm, slightly raspy voice (S1) says:
<d>[English] ... </d>
The spoken line itself. Inside the tag goes only a language tag and the exact words — preserved verbatim, never translated or rewritten. Delivery and action stay outside.
<d>[English] First batch of the morning.</d>
says in an off-screen voiceover
The exact phrase for voiceover. MiniMax also asks you to state, immediately after the line, that the on-screen character's lips stay closed.
The man (S1) says in an off-screen voiceover: <d>[English] I still remember that road.</d> while his lips remain completely closed.
<scenetrans>
Marks both sides of a line that carries across a cut, so speech is not restarted in the new shot.
his final words continue seamlessly across the cut
<cutoff>
Marks speech that is truncated by the end of the video rather than finished.
the sentence is clipped by the end of the clip
"..." on-screen text
Any sign, banner, subtitle or UI label that is actually visible goes in double quotes, verbatim. Describing text instead of quoting it is what produces letter-shaped noise.
A red neon sign reading "OPEN LATE" glows above the doorway.
N/A
The literal value for an empty audio field. Use it in non_diegetic_music when there is no score, and in overall_soundscape only when total silence is genuinely intended.
non_diegetic_music: N/A

Source: MiniMaxAI/MiniMax-H3 · docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md

H3Video method

The four-slot MiniMax H3 prompt structure

The official fields tell you what to fill in; they don’t tell you how to think about a shot. This is the working order we use inside integrated_multimodal_description — our simplification of the spec above, not a competing standard. Lighting, style and final-frame notes come after these four read clearly, not before.

  1. 01

    SUBJECT

    Who or what, with enough specificity to lock the appearance

    A silver-haired violinist in a rain-soaked wool coat

  2. 02

    ACTION

    One primary action, with an end state a viewer could point at

    lowers the bow and turns toward the doorway

  3. 03

    CAMERA

    The movement between frames — not a description of either one

    slow push-in from medium shot to close-up

  4. 04

    AUDIO

    Leave this out and the model picks a soundtrack you didn't choose

    solo violin fading into rain on glass, centered

H3Video method

The audio block of a MiniMax H3 prompt, in four parts

Picture and sound are generated together. Leaving this block out doesn’t give you a silent clip — it gives you a soundtrack the model chose. Silence has to be requested explicitly.

DIALOGUE
The lines themselves, plus tone and pacing cues
SOUNDSCAPE
Ambient sound and the music bed under it
STEREO PLACEMENT
Left, centre or right — H3 outputs native stereo, so use it
TIMING
The second a sound enters or drops out

MiniMax's own spec

The camera moves MiniMax H3 recognises

“Use camera language” is advice you can find in any MiniMax H3 prompt guide. These are the specific terms MiniMax documents, which is the part nobody reprints.

  • Zoom In / Zoom Out
  • Push In / Push Out
  • Pan
  • Truck
  • Tilt
  • Pedestal
  • Arc Shot
  • Tracking Shot
  • Static Shot
  • Shake
  • POV
  • Roll

How to write them

Each takes an amplitude (small or large) and a speed (slow or fast). MiniMax asks for these written as a natural English action within the shot rather than stacked as separate labels.

Source: MiniMaxAI/MiniMax-H3 · docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md

H3Video method

Six MiniMax H3 prompt rules that change the output

Schema and notation get a MiniMax H3 prompt parsed correctly. These six are about what to put inside it — the decisions that show up in the finished clip rather than in the parse.

  • Write the video as timed beats, not as a description

    Anything longer than a single action needs an order. Give each beat one primary change and an observable end state — something a viewer could point at.

  • Put the most important beat in the middle, not at the end

    The final beat is the one most likely to get squeezed. A climax parked at the end often arrives as half a movement.

  • Budget roughly four seconds for a prop change or hand-off

    Five complex actions in a fifteen-second clip means five unfinished actions. Fewer beats with enough room beats more beats rushed.

  • If only distance or angle changes, move the camera — don't cut

    A cut should introduce new information: subject, space, state, viewpoint or time. Cutting just to change angle makes the clip feel scattered.

  • Silence has to be requested explicitly

    Picture and sound are generated together. An empty audio block is not silence — it is a soundtrack chosen for you.

  • With first and last frames, describe the motion between them

    The two images already establish appearance. Restating it wastes the prompt; what it needs is the transition — the most precise control H3 offers.

H3Video library · 24 prompts

The MiniMax H3 prompt library — 24 prompts

This is the part of the MiniMax H3 prompt guide worth bookmarking. Every MiniMax H3 prompt below is written in the three-field format above, across 8 categories. Copy one, or load it into the generator and change a single line so you can tell what actually moved.

These are ours, written to MiniMax’s documented schema — we have not published renders of them, so treat them as starting points rather than guaranteed results. The library grows as we work through the model; the notation, not the subject matter, is what each one is teaching.

T2VText to Video

Rain, violin, doorway

Cinematic

DemonstratesThe plain three-field shape, with camera motion written as an action

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a medium-wide shot frames a silver-haired violinist in a rain-soaked wool coat standing under a shop awning at night. The camera pushes in with small amplitude at slow speed as she finishes a phrase, lowers the bow, and turns toward the doorway behind her, ending on her face in close-up.

overall_soundscape

Rain drums steadily on the canvas awning above a low street ambience. Water runs from the edge of the fabric onto the pavement, and a door hinge creaks once behind her.

non_diegetic_music

A solo violin line at a slow tempo, warm and slightly detuned, thinning to a single sustained note before it fades.

T2VText to Video

Night market hand-off

Cinematic

DemonstratesA second shot with an explicit cut timestamp

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a wide static shot frames a night market stall selling grilled squid, steam rising through orange string lights. The vendor fans the coals until embers flare and settle, then lifts a skewer and turns it once. [Shot 2] At 00:06.000, the shot cuts to a tracking shot that follows the skewer across the counter into a waiting customer's hand, then holds as the customer walks out of frame and leaves the stall empty and still smoking.

overall_soundscape

Fat sizzles over hot coals at the centre of the frame while a broad crowd murmur spreads across the street behind it. A bicycle bell passes from left to right, followed by the scrape of tongs on the grill.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Desert road, one take

Cinematic

DemonstratesHolding a single shot instead of cutting for variety

integrated_multimodal_description

[Shot 1] Live-action, cinematic, vintage film grain, a low static shot frames a two-lane desert road at midday with heat shimmer across the far tarmac. A single dust-covered pickup enters from the far distance and grows in the frame; the camera holds a static shot throughout, letting the truck pass close to the lens and exit right while the empty road settles back into shimmer.

overall_soundscape

A constant dry wind moves grit across the tarmac. The engine builds from a distant drone into a full pass with tyre roar, then falls away as the wind returns alone.

non_diegetic_music

Sparse baritone-guitar notes at a slow tempo with long gaps between them, ending on a single unresolved note.

T2VText to Video

Two-hander at a diner counter

Dialogue & voice

DemonstratesTwo stable speaker IDs alternating inside one shot

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a medium two-shot frames a night-shift waitress and a tired customer across a chipped diner counter. The camera holds a static shot as the waitress with a dry, low voice (S1) refills his cup and says: <d>[English] You want the pie or the story that comes with it?</d> The customer, a hoarse man in his sixties (S2), turns the cup a quarter-turn and answers: <d>[English] Whichever one takes longer.</d> She sets the pot down and lets the pause sit.

overall_soundscape

A refrigeration unit hums beneath the room while coffee pours into a ceramic cup. A spoon touches the saucer once, and traffic passes faintly beyond the window glass.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Voiceover over a night drive

Dialogue & voice

DemonstratesThe exact voiceover phrasing, plus the lips-closed statement it requires

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a close side shot frames a man driving at night, dashboard light on one side of his face and passing sodium lamps on the other. The camera trucks right with small amplitude at slow speed. The man (S1) says in an off-screen voiceover: <d>[English] I still remember that road.</d> while his lips remain completely closed. He checks the mirror and the lamps continue to sweep across the windscreen.

overall_soundscape

Tyre roar and a steady ventilation hum fill the cabin. The indicator ticks twice, and a passing truck sweeps from right to left.

non_diegetic_music

Sustained low strings at a slow tempo with widely spaced piano tones, gradually decreasing in volume.

T2VText to Video

A line that survives the cut

Dialogue & voice

DemonstratesCarrying one sentence across a shot change with <scenetrans>

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a close shot frames a young engineer in a server room, blue indicator light across her face. She begins mid-sentence with a clipped, precise voice (S1): <d>[English] If the rack drops again tonight</d> <scenetrans> [Shot 2] At 00:04.000, the shot cuts to a wide shot of the empty control room as her voice continues uninterrupted into the next shot: <scenetrans> <d>[English] nobody is going to be here to catch it.</d> The camera pans right with small amplitude at slow speed across the unattended desks.

overall_soundscape

Server fans hold a constant broadband roar that drops to a duller hum in the control room. A door seal releases once during the change of space.

non_diegetic_music

A low electronic pulse at a slow tempo that continues unchanged across the cut.

I2VImage to Video

Product orbit from a still

Product

DemonstratesA single clean camera idea plus an explicit silence request

Alignment instruction · first line

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

integrated_multimodal_description

[Shot 1] Live-action, cinematic, the product shown in <Picture 1> stays exactly where it is on the plinth, preserving its position, finish and the background gradient. The camera arcs left with large amplitude at slow speed through a half circle with the horizon locked, ending on the reverse side. The key light stays fixed, so highlights sweep across the surface as the camera travels.

overall_soundscape· left silent on purpose

N/A

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Macro pour, foley-led

Product

DemonstratesLetting the soundscape carry the clip instead of a score

integrated_multimodal_description

[Shot 1] Live-action, cinematic, an extreme close macro shot frames a chilled green glass bottle against a dark backdrop, condensation beading on the shoulder. The camera pushes in with small amplitude at slow speed as a hand tilts the bottle and a clear stream falls into a heavy tumbler, ice shifting once as the level rises. The pour stops and the last drop breaks the surface.

overall_soundscape

A cap breaks its seal with a short hiss. Liquid falls into glass with a rising pitch as the tumbler fills, ice cracks once, and a final drop lands cleanly before the room falls quiet.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Hands-only unboxing

Product

DemonstratesA top-down static frame with no face and no music

integrated_multimodal_description

[Shot 1] Live-action, cinematic, an overhead static shot frames a matte grey box centred on a pale oak desk with two hands entering from the bottom of the frame. The hands lift the lid straight up, set it aside outside the frame, fold back a sheet of tissue, and turn the product a quarter-turn so its face reads to the lens. The camera holds a static shot for the full clip and the composition never changes.

overall_soundscape

Card slides against card as the lid lifts, tissue crackles softly, and the product sets down on the desk with a single dull contact. Room tone stays low and even throughout.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Paper cut-out town

Animation

DemonstratesA style contract stated in the first line of Shot 1

integrated_multimodal_description

[Shot 1] 2D-animated paper cut-out style with visible fibre edges and flat layered depth, a side-on static shot frames a small harbour town at dawn. Layers of card slide horizontally at different speeds as a cut-out ferry crosses the bay from left to right; a paper gull hinges its wings twice above the roofline. The camera trucks right with small amplitude at slow speed to keep the ferry centred.

overall_soundscape

Soft paper rustles accompany each layer that moves. Water laps in a gentle loop and a single flat bell note rings from the harbour.

non_diegetic_music

A plucked ukulele figure at a moderate tempo with light hand-percussion, ending on one held chord.

T2VText to Video

Claymation kitchen

Animation

DemonstratesAsking for stop-motion cadence rather than smooth motion

integrated_multimodal_description

[Shot 1] Claymation with visible fingerprints in the surface and a slightly stepped, stop-motion cadence rather than smooth interpolation, a medium static shot frames a clay cook at a tiny kitchen bench. He lifts a clay pot, tips three round vegetables into it, and stirs twice while his eyebrows rise on each stir. The camera tilts down with small amplitude at slow speed to settle on the pot.

overall_soundscape

Soft clay contacts thud against the bench, a wooden spoon knocks twice inside the pot, and a low bubbling loop rises underneath.

non_diegetic_music

A bassoon and pizzicato-string figure at a moderate tempo, punctuated by a woodblock on each stir.

T2VText to Video

Watercolour street in rain

Animation

DemonstratesStyle plus a transition described as an event, not an effect

integrated_multimodal_description

[Shot 1] Watercolour style with soft bleeding edges and visible paper grain, a wide static shot frames a narrow street where rain spreads across the washes. A figure with a red umbrella walks away from the lens down the centre of the street; the pigment of the buildings runs slightly downward as the rain falls, and the far end of the street blooms into pale paper as she reaches it.

overall_soundscape

Rain falls evenly on stone with occasional heavier drips from the gutters. Footsteps land in shallow water at a steady pace and fade with distance.

non_diegetic_music

A single piano at a slow tempo with wide spacing between notes, decreasing in volume as the street pales.

T2VText to Video

Rain on a window, no score

Audio-led

DemonstratesWriting overall_soundscape as the main event and refusing music

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a close static shot frames a single-glazed window from inside a dark room, with rain running down the outer surface and a blurred street beyond. Droplets merge and accelerate down the glass; the camera holds a static shot for the entire clip and nothing else in the frame moves.

overall_soundscape

Rain taps unevenly against the glass at close range, with heavier drops striking the sill below. A distant gutter runs continuously and one car passes on wet tarmac far away.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Stairwell footsteps

Audio-led

DemonstratesPhysical action sound placed in the soundscape field

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a high-angle static shot looks straight down a concrete stairwell as a figure descends past the camera through four flights. The camera holds a static shot and the figure grows smaller with each turn of the stair.

overall_soundscape

Hard-soled footsteps strike concrete with a long reverb tail that shortens as the figure descends. A hand slides intermittently along a metal rail, fabric brushes at each turn, and breathing stays audible under the steps.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Rehearsal room, diegetic vs score

Audio-led

DemonstratesKeeping music the characters can hear out of non_diegetic_music

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a medium-wide shot frames three musicians in a carpeted rehearsal room under a single work lamp. The drummer counts in with sticks, and the trio plays a slow, loose blues figure together — this music is played in the room and is audible to the characters. The camera arcs right with small amplitude at slow speed around the kit as the guitarist leans into the first turnaround.

overall_soundscape

Room reflections are short and dead from the carpet. Amplifier hiss sits under the performance, a chair leg scrapes once, and a stick taps the rim between phrases.

non_diegetic_music· left silent on purpose

N/A

R2VReference to Video

Character kept across a new scene

References

DemonstratesNaming each reference and giving it exactly one job

integrated_multimodal_description

[Shot 1] Live-action, cinematic, the woman in <Picture 1> keeps her face, hair length and olive field jacket exactly as shown; <Picture 2> supplies the rooftop location and its blue-hour light, and nothing else. She stands at the parapet with wind moving her hair, checks her watch, exhales, and lifts her gaze as an aircraft crosses above. The camera holds a low-angle medium shot and tilts up with small amplitude at slow speed to follow her look.

overall_soundscape

Wind moves across the microphone in irregular gusts over a low city hum. A jacket collar snaps once, and a distant jet builds and passes overhead.

non_diegetic_music

Sustained synth pads at a slow tempo with a single low pulse, increasing in volume as the aircraft passes.

R2VReference to Video

Motion from a clip, look from stills

References

DemonstratesSplitting the job between a video reference and image refs

integrated_multimodal_description

[Shot 1] Take the camera path and the timing of the movement from the reference video, and take the palette, grain and lighting from <Picture 1> and <Picture 2> only. A paper crane sits centred on a dark table and unfolds itself in reverse into a flat sheet, each crease releasing in sequence. The camera holds an overhead static shot with the sheet centred throughout.

overall_soundscape

Dry paper creases release one at a time, sharp and close to the lens, with a final soft settle as the sheet lies flat. Room tone is minimal.

non_diegetic_music· left silent on purpose

N/A

R2VReference to Video

Voice taken from an audio reference

References

DemonstratesBinding a reference voice to a specific speaker ID

integrated_multimodal_description

[Shot 1] Live-action, cinematic, the man shown in <Picture 1> keeps his face, beard and grey overcoat; the reference audio supplies his voice timbre and speaking rate for (S1) and nothing else. He stands at a bus stop in early light, checks the timetable, and turns to the lens. The man (S1) says: <d>[English] It never comes on time, but it comes.</d> The camera holds a static medium shot.

overall_soundscape

Early traffic passes at a distance over a low wind. A timetable panel rattles once in its frame and footsteps approach from off-screen.

non_diegetic_music· left silent on purpose

N/A

I2VImage to Video

Portrait turn between two frames

First & last frame

DemonstratesDescribing only the path, since both frames are given

Alignment instruction · first line

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 5.17-second mark of the target video.

integrated_multimodal_description

[Shot 1] Live-action, cinematic, the subject begins in the profile position and framing established by Picture 1. The camera holds a static shot as she rotates her head from profile toward the lens, her scarf lifting slightly with the motion and settling again, until her eye line, head angle and the scarf's fall match the composition established by Picture 2 at the end of the shot.

overall_soundscape

Room tone is very quiet and even. Fabric rustles once as the scarf lifts, and a single breath is audible near the end.

non_diegetic_music· left silent on purpose

N/A

I2VImage to Video

Umbrella opening in the rain

First & last frame

DemonstratesNarrowing differences until the last frame is reached

Alignment instruction · first line

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot 1) aligns with the 8.00-second mark of the target video.

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a rain-soaked cyclist begins in the position and framing established by Picture 1, holding a closed umbrella beside her bicycle. The camera pulls out with small amplitude at slow speed as she releases the handlebar, raises the umbrella above her shoulder, and presses the runner upward until the canopy opens. Water rolls off the expanding fabric while she steps beneath it and rotates the handle into the final angle, settling into the pose, spacing and composition established by Picture 2 at the end of the shot.

overall_soundscape

Rain falls steadily on pavement, followed by the metallic click of the umbrella runner and the soft snap of the canopy opening. Water drips from the bicycle frame as distant traffic passes.

non_diegetic_music· left silent on purpose

N/A

I2VImage to Video

Landing on a last frame

First & last frame

DemonstratesInferring a plausible earlier state and converging onto it

Alignment instruction · first line

How the reference pictures align with the target video — <Picture 1> (from [Shot 1]) aligns with the 6.58-second mark of the target video.

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a close shot begins with an intact drinking glass near the edge of a dark wooden table while the same hand and sleeve visible in <Picture 1> approach from the right. The camera pushes in with small amplitude at slow speed as the fingertips catch the rim. The glass tips, falls and strikes the floor; cracks spread and fragments slide outward, then lose momentum and settle into the exact broken arrangement, hand position, camera angle and lighting established by <Picture 1>.

overall_soundscape

Fingertips tap the glass before it scrapes across the tabletop, falls and breaks with a sharp crash. Small fragments scatter and gradually stop sliding across the floor.

non_diegetic_music

A low electronic pulse at a slow tempo, ending immediately after the glass breaks.

T2VText to Video

Neon sign, spelled out

On-screen text

DemonstratesQuoting the exact string instead of describing the sign

integrated_multimodal_description

[Shot 1] Live-action, cinematic, a low-angle static shot frames the frontage of a late-night noodle bar in the rain. A red neon sign reading "OPEN LATE" glows above the doorway and a smaller window decal reads "NO. 42". The sign's second word flickers twice and steadies while steam crosses in front of it. Do not add any other text to the frame and do not misspell either string.

overall_soundscape

Rain falls on an awning and a metal extractor fan hums above the door. The neon transformer buzzes faintly and a door opens once, releasing kitchen noise onto the street.

non_diegetic_music· left silent on purpose

N/A

T2VText to Video

Interface walkthrough

On-screen text

DemonstratesVerbs for interactions and every label written out

integrated_multimodal_description

[Shot 1] 3D CG, a clean screen-recording look, a static shot frames a dark dashboard interface. A cursor moves to a button reading "Generate", presses it, and a progress bar fills from left to right beneath a header reading "Render queue". A card then expands to reveal a label reading "Ready". Every visible string is exactly as quoted; do not add other text and do not misspell any word.

overall_soundscape

A soft click accompanies the press, a short ascending tone marks completion, and a low interface hum runs underneath.

non_diegetic_music

A minimal electronic pulse at a moderate tempo with a single soft chord on completion.

T2VText to Video

Title card with an audio sting

On-screen text

DemonstratesTiming a sound to a visual beat inside one shot

integrated_multimodal_description

[Shot 1] Live-action, cinematic, vintage film grain, a static shot on black. Fine dust drifts through a narrow beam of light, then a title reading "NORTH OF HERE" resolves at the centre of the frame, holds, and dims. The camera holds a static shot throughout and the title is the only text in the frame.

overall_soundscape

A projector-like flutter runs quietly under the black frame, and a single low impact lands as the title resolves. The flutter continues after the title dims.

non_diegetic_music

A sustained low brass note at a slow tempo that swells with the title and cuts off as it dims.

H3Video method

How to iterate a MiniMax H3 prompt without burning generations

A MiniMax H3 prompt has four things that can each be wrong independently — framing, motion, dialogue and sound. Changing all four between runs tells you nothing. Change them in this order instead:

  1. 01

    Framing first, at the shortest duration

    Get the subject, style contract and opening composition right before anything moves. A wrong style contract poisons every later pass.

  2. 02

    Then one camera move

    Add a single motion with its amplitude and speed. If the frame drifts, the amplitude is too large, not the prompt too short.

  3. 03

    Then the cut, if you need one

    Add [Shot 2] with an explicit timestamp. If the second shot adds no new information, delete it and use camera motion instead.

  4. 04

    Then dialogue, verbatim

    Add the speaker ID and the <d> tag last. Lines change lip motion and pacing, so adding them earlier makes every earlier comparison invalid.

  5. 05

    Sound at every stage, never empty

    Keep overall_soundscape and non_diegetic_music filled — or set to N/A — from the first pass. An empty field is not neutral.

One line, twice

Test one controlled prompt change

No MiniMax H3 prompt guide can tell you what one line change does to your shot. Compare the same MiniMax H3 prompt twice with one line different, then review the completed clips side by side.

Open pricing →

Paid text-to-video uses the published per-second credit rate.

Prompt iteration is cheap in wall-clock terms only if generation is cheap. If you’re running locally, check the VRAM requirements before committing to a long clip — a failed 15-second generation costs you the whole run. The ComfyUI workflow guide covers which template accepts which inputs, which matters here: the reference-driven workflow takes up to nine images, and it also explains why the duration you type is not the duration you get — worth knowing before you write timestamps into a shot list.

If you are iterating a lot, the turbo LoRA cuts sampling from 20 steps to a handful, which changes prompt iteration from a coffee break into a loop you can actually stay inside.

Frequently asked questions

How do you structure a MiniMax H3 prompt?

Use SUBJECT → ACTION → CAMERA → AUDIO. Name who or what, give one primary action with an observable end state, describe the camera as movement rather than a still description, then write an explicit audio block. Add lighting and style only after those four read clearly.

Does MiniMax H3 generate audio automatically?

Yes, and that is the problem. Picture and sound are generated together, so leaving the audio block empty does not produce silence — it lets the model pick a soundtrack you did not choose. Silence must be requested explicitly.

How do I control the camera in MiniMax H3?

Use explicit camera language — push-in, tracking shot, pan, tilt, orbit, static close-up — and describe the motion between frames rather than what any single frame looks like. If only the distance or angle needs to change, use camera motion instead of a cut.

Why does the end of my video feel rushed?

The final beat is the one most likely to get squeezed. Put your most important beat in the middle of the timeline instead of the end, and budget roughly four seconds for anything involving a prop change or hand-off.

Is there an official MiniMax H3 prompt guide?

Yes. MiniMax ships docs/VIDEO_PROMPT_WRITING_GUIDE_base_en.md with the model on Hugging Face. It defines three required fields — integrated_multimodal_description, overall_soundscape and non_diegetic_music — plus notation for shots, speakers and dialogue. Every prompt in the library on this page is written in that format.

How do I write dialogue in a MiniMax H3 prompt?

Give the speaker a stable ID such as (S1), establish their voice and delivery outside the tag, and put only a language tag and the exact words inside a <d> tag: <d>[English] I get off at the next station.</d>. Preserve punctuation verbatim. For voiceover, use the phrase "says in an off-screen voiceover" and then state that the character's lips remain closed.

Why does on-screen text render as gibberish?

Because it was described instead of quoted. Any sign, subtitle or UI label that should be readable has to be spelled out in double quotation marks exactly as it should appear, and it helps to add that no other text should be added and nothing should be misspelled.

How long should a MiniMax H3 prompt be?

Long enough to fill all three fields and no longer. One shot with one clear camera move, a soundscape of one to four sentences, and one to three sentences of score is usually enough. Extra shots cost duration you have not budgeted, and the last beat is the one that gets squeezed.