Prompting guide
How to prompt MiniMax-H3
How to prompt MiniMax-H3: its labeled prompt format, a fill-in template, and expert tips for shots, frames, references, dialogue, and sound on Mage.
Quick start
Your first great MiniMax-H3 prompt
Everything a beginner needs in one minute: what to know first, the rules that matter most, a template to fill in, and prompts to copy.
What to know first
MiniMax-H3 makes video and sound together, and on Mage it reads your prompt exactly as you write it: nothing expands or reorganizes it first. It was trained on prompts written in labeled fields, so write in those fields instead of a loose paragraph. For a clip from text, three fields are enough: what we see and hear, shot by shot; the ambient sound of the whole clip; and the music, or N/A for none.
The rules that matter most
Write the three fields, in order
integrated_multimodal_description, then overall_soundscape, then non_diegetic_music, each on its own line with a blank line between. Copy the labels exactly.
Open each shot with its style and framing
[Shot 1] starts with the style and the starting picture: "Live-action, cinematic, a medium shot frames…". Then one camera move as a sentence, then what happens.
Fill the length exactly
The shots must cover the clip length you set, no more and no less. A cut is written with its time: "[Shot 2] At 00:04.000, the camera cuts to…".
Keep sound in its place
Sounds tied to an action go in the description; the room, the weather, and the crowd go in overall_soundscape; score goes in non_diegetic_music, or N/A for none.
Write sections in English, words in their own language
The fields are in English. Dialogue, lyrics, and on-screen text stay in their original language, exactly as they should appear.
Fill in this template
Replace each part in square brackets with your own idea, and leave out a part you do not need.
integrated_multimodal_description: [Shot 1] [Style], a [shot size] frames [subject, place, and starting state]. The camera [one move, as a sentence]. [What happens, in order, with the sounds of each action].
overall_soundscape: [Ambience and physical sounds across the whole clip].
non_diegetic_music: [Score instruments, tempo, and dynamics, or N/A].integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium shot frames a noodle cook at a street stall at night, steam rising from a row of pots under a single bare bulb. The camera pushes in with small amplitude at slow speed. He lifts a tangle of noodles high out of the boiling water, shakes them twice so drops scatter, and drops them into a bowl with a soft splash.
overall_soundscape: Rain patters on the tin awning, water boils steadily, and scooters pass on the wet street behind him.
non_diegetic_music: N/A[Shot 1]Keep this label exactly as written: it opens the first shot. A cut adds [Shot 2] with its time.[Style]Live-action, cinematic; 2D-animated; 3D CG; claymation; watercolor; vintage film.[shot size]How close the camera is: wide shot, medium shot, close-up.[subject, place, and starting state]Who or what we see, where, and how the shot begins: a noodle cook at a street stall at night, steam rising from a row of pots.[one move, as a sentence]One camera move, with its size and speed when they matter: pushes in with small amplitude at slow speed.[What happens, in order, with the sounds of each action]The action as it unfolds, and the sound each action makes: he lifts the noodles high and drops them into a bowl with a soft splash.[Ambience and physical sounds across the whole clip]One to four sentences of what we hear throughout: rain on the awning, boiling water, passing scooters.[Score instruments, tempo, and dynamics, or N/A]Music only the audience hears, as instruments, tempo, rhythm, and loudness, with no mood words. N/A for no music.
Prompts to start from
One shot, sound but no music
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a close-up frames a glassblower's hands turning a glowing orange bulb of glass on the end of a steel pipe in a dark workshop. The camera trucks left with small amplitude at slow speed. She rolls the pipe along the rails of her bench, lifts it, and blows a short breath into the mouthpiece as the bulb swells and brightens. overall_soundscape: The furnace roars steadily behind her, the pipe clicks against the rails, and a soft hiss rises as the glass swells. non_diegetic_music: N/AWhy it works: Three fields in order, one camera move written as a sentence, and N/A for music, so the clip sounds like the workshop and nothing else.
A cut and a spoken line
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames two hikers reaching a rocky summit above a sea of clouds at sunrise, wind tugging at their jackets. The camera pedestals up with small amplitude at slow speed as they climb the last step and stop. [Shot 2] At 00:04.000, the camera cuts to a close-up of the woman in the red jacket (S1), in her thirties, with a bright, breathless voice, who turns to her friend and says: <d>[English] Told you it was worth the early alarm.</d> Her friend laughs and shakes his head. overall_soundscape: Strong wind gusts across the summit, boots scrape on rock, and both hikers breathe hard from the climb. non_diegetic_music: A sustained string pad at a slow tempo, quiet at first and rising gently after the cut.Why it works: The cut at 4 seconds brings something new, a face and a line; the speaker gets an ID and a voice description once; and only the words sit inside the d tags. Set the length to 8 seconds.
Starting on your image
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 old bookshop shown in <Picture 1> keeps its layout, warm lamps and crowded shelves, with the gray cat asleep on the counter. The camera pushes in with small amplitude at slow speed. The cat opens one eye, stretches its front paws across an open ledger, and yawns as a page lifts in the draft from the door. overall_soundscape: A wall clock ticks, rain murmurs against the front window, and the floorboards creak once. non_diegetic_music: A solo piano plays a sparse melody at a slow tempo and low volume.Why it works: Upload the still as the First Frame. The first line is copied exactly, the description opens on what the image already shows, and only then says what moves.
From one image to another
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 6.00-second mark of the target video. integrated_multimodal_description: [Shot 1] Live-action, cinematic, an empty sunlit studio begins as in Picture 1, a folded wooden easel leaning against the wall. The camera holds a static shot. A painter in a paint-flecked apron walks in from the right, unfolds the easel, sets a blank canvas on it, and steps back with her arms crossed, settling into the pose and spacing of Picture 2. overall_soundscape: Footsteps cross the wooden floor, the easel legs creak and lock with a clack, and birdsong drifts through an open window. non_diegetic_music: N/AWhy it works: Upload the two stills as the First Frame and the Last Frame. The first line pins them to the start and the end, and the description tells the path between them instead of describing them again. Set the length to 6 seconds.
Advanced
Advanced MiniMax-H3 prompting
Techniques for more control once the basics work.
MiniMax-H3 Turbo or MiniMax-H3
MiniMax-H3 Turbo is the default: faster, cheaper, and the one for longer clips. MiniMax-H3 is the full-step model, for shorter clips. Both read a prompt the same way, so draft on Turbo and render on either with the same words.
- Leave Prompt Enhance off. It adds words to the end of the prompt, where they land in the music field.
The format, field by field
MiniMax-H3 reads one of two layouts. For a clip from text or from frames, write an optional frame line, a blank line, then three fields: integrated_multimodal_description, overall_soundscape, and non_diegetic_music. For a clip from reference images, Characters, or reference videos, write six: subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape, and non_diegetic_music.
integrated_multimodal_description is the body of the clip: a timeline of shots, each with its picture, its camera, its actions, and any dialogue or singing. overall_soundscape takes one to four sentences of ambience, physical sounds, breaths, and laughter across the whole clip. non_diegetic_music takes one to three sentences of score, or N/A.
| Write | What it does |
|---|---|
integrated_multimodal_description: | Shot by shot: the style, the framing, the camera, the action, and any speech. |
overall_soundscape: | Ambience and physical sounds through the whole clip. N/A only for total silence. |
non_diegetic_music: | Score the characters cannot hear, or N/A for none. |
subject_definitions: | With references: what each label is, defined once before it is used. |
summary: | With references: the kind of task in brackets, then one short paragraph on the clip. |
retention_analysis: | With references: how much of each input to keep. |
detailed_description: | With references: one or two style sentences, then the shots. |
Shots and timing
[Shot 1] has no time: it opens with the style and the starting picture ("Live-action, cinematic, a medium-wide shot frames…"). Every later shot starts with its time in minutes, seconds, and milliseconds: "[Shot 2] At 00:05.000, the camera cuts to…". Times must rise from shot to shot and stay inside the clip, and the last shot runs to the end.
Every cut must add something new: a new subject, a new space, a new state, a new viewpoint, or a new time. For a tighter framing on the same moment, move the camera instead of cutting.
- Write cuts as "the camera cuts to" or "the shot switches to". Ask for a cross-dissolve, a fade, or a wipe only when you want one.
- At every cut, say the framing and who is on screen.
- For one unbroken take, write a single [Shot 1] and say "in one continuous shot with no cuts". Never ask for a continuous take and also write several shots.
- Style words for the start of [Shot 1]: Live-action, cinematic; 2D-animated; 3D CG; claymation; watercolor; vintage film. With frames, take the style from the image.
- Write concrete detail, not praise: "shallow depth of field, backlight keeps the foreground dark", never "beautiful, cinematic".
Camera moves
Write each camera move as a sentence inside its shot, not as a tag at the end: "The camera holds a static shot as the runner exits the frame." Add a size and a speed only when they matter: with small or large amplitude, at slow or fast speed.
- For an orbit around a subject, write a truck and a pan in opposite directions: "The camera trucks left while panning right."
| Write | What it does |
|---|---|
push in / pull out | The camera travels toward or away from the subject. |
zoom in / zoom out | Only the lens changes; the camera stays put. |
pan left / pan right | The camera turns on the spot, left or right. |
truck left / truck right | The camera slides sideways. |
tilt up / tilt down | The camera tips up or down on the spot. |
pedestal up / pedestal down | The whole camera rises or lowers. |
arc shot | The camera curves around the subject. |
tracking shot | The camera travels with a moving subject. |
static shot | The camera does not move. |
shake slightly / shake strongly | Handheld or impact shake. |
POV | The view through a character's eyes. |
roll clockwise / roll counterclockwise | The frame rotates. |
with small amplitude at slow speed | A gentle, slow version of any move. |
First and last frames
Upload a First Frame, a Last Frame, or both, and open the prompt with the matching line, copied exactly, then a blank line and the three fields. In the line, S.SS is the clip length with two decimals (6.00 for six seconds), and N is the number of the last shot. The clip takes the shape of the frame.
With a first frame, open the description on what the image shows (its style, subjects, and composition), keep them, then write how the action starts, develops, and ends. With both frames, write one shot unless you want cuts, and describe the path between them: the first state, the visible changes, the narrowing differences, and the last state. With only a last frame, invent a believable earlier state and arrive at the image in the final shot.
- Never describe the two stills again in a first-and-last clip: the model already has them.
- One clip uses frames or references, not both.
| Write | What it does |
|---|---|
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. | First frame only. |
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 N) aligns with the S.SS-second mark of the target video. | First and last frames. |
How the reference pictures align with the target video — <Picture 1> (from [Shot N]) aligns with the S.SS-second mark of the target video. | Last frame only. |
References and Characters
MiniMax-H3 numbers its inputs by type, and these labels are what it reads. <Picture N> counts your uploaded images first (the Image, then the Additional Images), then the Characters you mention, then the References you mention, each in the order of first mention. <Audio N> counts the voice clips you mention, then the voices of mentioned Characters, then the sound of each reference video. <Video N> counts your Video References in order.
Before the model reads the prompt, Mage turns each @handle into words such as "character from image 3 with voice from [Audio 1]". So mention each handle once, in subject_definitions beside its label, and use only the labels after that. An uploaded image needs no handle: write its <Picture N> label directly.
A <Subject N> is something to reuse: a person, a place, an object, an outfit, a style, or a motion. One subject can draw on several inputs, such as "whose face comes from <Picture 1> and whose walk comes from <Video 1>". Use <Picture N> on its own only when the image is a frame or a storyboard for given shots.
- Give every input one stated role: character, object, setting, frame, voice, storyboard, style, composition, motion, camera motion, or audio to reuse. Name the features that must not change.
- A face holds only when the person is attached, through a Character or an image. Text alone never keeps a face.
- Keep each label for the whole prompt, and never define it twice or give one input two labels.
- The summary opens with the kind of task in brackets, joined by +: keyframe completion, reference generation, video editing, video continuation, audio reuse, or audio reference.
- To edit a clip, add it as a Video Reference and open the summary with [video editing] and "The target video is an edited version of <Video 1>."
- References share a time budget with the clip: each image and each second of reference video uses part of it. Trim reference videos to the part you need, and keep references few on long clips.
- Write a detailed_description of about 350 to 500 words for a new clip. For each shot, give the composition, who stands where, the setting and light, the actions, the camera, the sound, and where each label applies.
| Write | What it does |
|---|---|
fully_preserved | Keep it exactly: a face, a product, a logo. |
partially_preserved | Keep some of it, and say which parts. |
attribute_transfer | Take one quality, such as a color, a texture, or a style. |
weak_reference | A loose guide only. |
fully_copy, partially_copy, reference, weak_reference | The same scale for audio: play it as it is, play part of it, follow only its timbre or rhythm, or treat it as a loose guide. |
A saved Character with a voice, and a product photo
subject_definitions:
<Subject 1> is the man in <Picture 2> (@leo), with a gray beard, a flat cap and a navy wool coat.
<Subject 2> is the cast-iron teapot in <Picture 1>, matte black with a bamboo handle.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1).
summary:
[reference generation + audio reference] <Subject 1> pours tea from <Subject 2> at a table by a rainy window and speaks to the camera in the voice referenced from <Audio 1>.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - face, beard, cap and coat are kept.
<Subject 2> (appears in [Shot 1], [Shot 2]): fully_preserved - shape, finish and bamboo handle are kept exactly.
<Audio 1>: reference - his delivery follows its timbre without copying its words.
detailed_description:
Live-action, soft gray daylight from a window on the left, shallow depth of field.
[Shot 1] A medium shot frames <Subject 1> seated at a worn oak table, lifting <Subject 2> and pouring a thin stream of tea into a small white cup as steam curls up. The camera pushes in with small amplitude at slow speed. [Shot 2] At 00:04.000, the camera cuts to a close-up of <Subject 1> (S1), who looks into the lens and says in a slow, warm voice: <d>[English] Rain outside, tea inside. That's the whole plan.</d> He sets <Subject 2> down beside the cup.
overall_soundscape:
Rain runs down the window, tea splashes softly into the cup, and the iron lid clinks as it settles.
non_diegetic_music:
N/AWhy it works: Upload the teapot as the Image, so it is <Picture 1> and the Character is <Picture 2>; the handle appears once, beside its label. Shortened here: a full description runs longer. Set the length to 8 seconds.
Dialogue and sound
Give every voice a stable ID, (S1), (S2), or (S1,S2) for a line spoken together, and describe each voice once, outside the tag, when it first appears: age, gender, on or off screen, pitch, timbre, pace, and accent. Characters who never make a sound get no ID.
Put each line inside <d> tags with only its language and the exact words: <d>[English] We leave at dawn.</d> Never translate a line, and fit the words to the shot: a long speech in a short shot breaks the lip sync.
Sort sound into three layers. Sounds tied to an action and music playing in the scene go in the description; ambience and physical sounds go in overall_soundscape; score goes in non_diegetic_music, written as instruments, tempo, rhythm, and dynamics without mood words. Never repeat dialogue in the two sound fields.
- Every clip has sound. Describe what you want to hear, and say what to leave out.
- With a voice used only for its timbre, leave its words out of the prompt. To reuse audio as it is, write out its exact words or lyrics, with [unclear] for any you cannot make out.
- Turn off Use Character Voices to leave Characters' saved voices out.
- Never write an email address or a social handle: any @ followed by letters is read as a saved handle. In dialogue, write "at".
| Write | What it does |
|---|---|
(S1) | The speaker ID, written right after the person who speaks. |
<d>[English] …</d> | A spoken line: the language, then the exact words. |
says in an off-screen voiceover: | Narration. Add "while her lips remain completely closed" after the line when the person is on screen. |
<scenetrans> | Placed at both sides of a cut when one line carries across it, with "continues seamlessly across the cut". |
<cutoff> | Speech that the end of the clip cuts off. |
<Audio 1> is the voice-timbre reference for <Subject 1> (S1). | A saved voice gives a speaker its sound; their words come from the prompt. |
A narrator over a wide shot
integrated_multimodal_description: [Shot 1] Live-action, cinematic, an extreme wide shot frames a herd of horses galloping across a frozen lake at dawn, snow spraying from their hooves. The camera trucks right with large amplitude at slow speed, keeping pace with the herd. A calm older woman (S1) with a low, clear voice says in an off-screen voiceover: <d>[English] Every winter, they cross before the ice turns.</d> No one on screen speaks.
overall_soundscape: Hooves drum on the ice, the horses snort and blow, and a cold wind hisses across the lake.
non_diegetic_music: Low cellos hold a long chord at a slow tempo, swelling slightly as the herd passes.Why it works: The narrator has an ID and a one-time description, the voiceover phrase keeps the line off screen, and each sound sits in its own layer.
Text on screen
Put any sign, label, subtitle, or neon in English double quotes, word for word, in its original language: A blue neon sign reading "NOODLES" flickers above the stall.
- Keep to one main piece of text per shot, and keep it off faces.
- When nothing should appear, write "No subtitles, on-screen text, or watermarks."
Common mistakes
Avoid this, do this instead
Avoid: A loose paragraph with no fields.
a cook makes noodles at a street stall at night in the rain, steam, cinematic, moodyInstead: The labeled fields: nothing on Mage expands a loose prompt.
integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium shot frames a noodle cook at a street stall at night, steam rising under a bare bulb. The camera holds a static shot. He lifts the noodles high and drops them into a bowl with a soft splash. overall_soundscape: Rain on the tin awning, boiling water, passing scooters. non_diegetic_music: N/AAvoid: A label used before it is defined, or one input under two labels.
Instead: Each label defined once in subject_definitions, then the same label every time.
Avoid: "One continuous take" with several shots.
Instead: One or the other.
Avoid: Music in one field and "no background music" in another.
Instead: One decision, in non_diegetic_music.
Avoid: A timeline longer or shorter than the clip.
Instead: Cut times inside the length you set, with the last shot running to the end.
Avoid: Mood words in the music field: "epic, emotional, uplifting".
non_diegetic_music: epic emotional uplifting musicInstead: Instruments, tempo, rhythm, and dynamics.
non_diegetic_music: Brass and timpani at a moderate tempo, building from quiet to loud over the last shot.Avoid: A long speech in a short shot.
Instead: A line short enough to say in the time the shot lasts.
Avoid: A plot summary: "they fall in love".
Instead: What is seen and heard, shot by shot.
Avoid: Expecting a face to hold from words alone.
Instead: The person attached as a Character or an image.
FAQ
MiniMax-H3 prompting questions
How long should a MiniMax-H3 prompt be?
Do I have to use the labeled fields?
How do I make a character speak in MiniMax-H3?
How do I keep a character consistent in MiniMax-H3?
How do I stop MiniMax-H3 from adding music?
Can I use first and last frames with references?
Does MiniMax-H3 support negative prompts?
Prompting guide
More prompting guides
How to prompt Berry
How to prompt Berry, Mage's reference-faithful video model: a one-minute quick start, a template, and expert tips for references, timing, dialogue, and edits.
How to prompt Blueberry
How to prompt Blueberry, Mage's first-and-last-frame video model: a quick start, a template, and expert tips for transitions, loops, camera moves, and sound.
How to prompt Cherry
How to prompt Cherry, Mage's video model: a one-minute quick start, a fill-in template, and expert tips for shots, camera, characters, sound, and edits.
How to prompt Grok Video
How to prompt Grok Video, xAI's Grok Imagine, on Mage: a quick start, a fill-in template, and expert tips for camera, references, edits, and sound.
How to prompt Kiwi
How to prompt Kiwi, Mage's everyday video model: a one-minute quick start, a fill-in template, and tips for light, camera moves, start images, and sound.
How to prompt Lemon
How to prompt Lemon, Mage's all-round video model: a one-minute quick start, a template, and expert tips for frames, references, voices, shots, and edits.
Make your first video with MiniMax-H3
New accounts get 300 free Gems, enough to try MiniMax-H3 Turbo. No install, no credit card.
Try MiniMax-H3 Turbo