The MiniMax H3 prompt guide in one screen
A MiniMax H3 prompt is a timeline with a sound track attached, not a paragraph of description. The model renders picture and stereo audio in the same pass, accepts up to twelve reference files, and gives you four to fifteen seconds of screen time. Write in five blocks, in this order: style contract, reference roles, timeline, camera, sound. Close with a short exclude list. Budget the seconds before you write a word.
That is the whole MiniMax H3 prompt guide, compressed. Below is the reasoning behind each block, a second-by-second budget, three copy-ready prompts, and a diagnostic that traces a defect in your clip back to the block that caused it. Treat it as a working MiniMax H3 prompt guide rather than a spec sheet.
What a MiniMax H3 prompt guide has to account for
If your habits come from image-first video models, three assumptions break at once.
Audio is not a post-production step. H3 generates dialogue, effects, and room tone alongside the frames. Silence in your prompt is not neutral; it invites the model to score the clip however it likes. Most complaints about unwanted background music trace back to a prompt that never mentioned sound.
References need jobs, not just uploads. You can attach nine images, three video clips, and three audio files, twelve in total, and audio never travels alone: it needs an image or video with it. A photo of a person in a kitchen contains a face, an outfit, a room, and a colour grade. Unless you say which of those the file controls, the model decides for you.
Multi-shot is native. H3 honours cuts inside one generation, which is why so many clips feel frantic: people request three shots in six seconds because nothing stops them.
Any MiniMax H3 prompt guide that skips those shifts hands you a formula that reads well and generates badly. Older MiniMax prompt guide advice, written for single-shot silent models, is the usual source of that mismatch.
The MiniMax H3 prompt guide starts with a second budget
Two numbers decide what fits. Natural speech runs at roughly two and a half words per second. A physical action worth seeing needs about two seconds to register. Multiply those against your duration and see which shot plans they forbid.
- Four to six seconds. One action, one camera move, ten spoken words at most. The reliable unit, and where to test every new idea.
- Seven to ten seconds. One action plus one reaction, one camera move, roughly twenty-two spoken words. Enough for a beat with a turn in it.
- Eleven to fifteen seconds. Three timeline blocks, one camera decision each, around thirty spoken words total, with silence at each end so the audio does not sound clipped.
This is the step most MiniMax H3 prompts skip, and it explains the commonest disappointment: a prompt that is well written, richly detailed, and impossible to perform in the time allowed. If your beat list exceeds the budget, cut a beat rather than compressing the ones you keep, because vagueness fails in a worse way. No MiniMax H3 prompt guide can rescue a shot plan that does not fit.
The prompt field holds seven thousand characters, far more than a fifteen-second clip needs. Long MiniMax H3 prompts are not better prompts. Length is not your constraint. Time is.
The five blocks of the MiniMax H3 prompt guide
Style contract
Two or three sentences fixing medium, palette, lighting, and era: shot on 35mm with fine grain, muted teal and sand palette, overcast midday light, late-1990s documentary feel. Treat it as a contract, because everything downstream inherits from it. Put it first, since a style block at the bottom reads as decoration and gets weighted like decoration.
Reference roles
Number every file and assign it exactly one job. The role map is the highest-leverage habit in this entire MiniMax H3 prompt guide, and it takes ten seconds to write.
Image 1: subject identity only, face and hair, ignore clothing. Image 2: wardrobe only. Image 3: location and colour grade only. Video 1: body motion and timing only, ignore its subject and setting. Audio 1: vocal timbre and delivery only, not the words spoken.
Two details save most reruns. Name three or four concrete anchors that must not change, and say when a reference should never appear on screen. A character sheet used as a style source will otherwise turn up as your opening frame.
Timeline
Write time-coded blocks, not a run-on paragraph. A shot list that collapses into prose is the failure mode behind most disappointing long prompts.
[0-3s] She sets the mug down and looks up. [3-7s] She stands, crosses to the window. [7-10s] She stops, exhales, holds still.
Chronological order, one clear action per block, present tense. For a continuous take, say so and delete every cut instruction: asking for one unbroken shot and three cuts produces neither. A MiniMax H3 prompt guide that stops at describe the action leaves this out, and it is where ambitious prompts come apart.
Camera
One camera idea per shot. That single constraint does more for a MiniMax H3 prompt guide than any vocabulary list, though the vocabulary reads directly: push in, pull out, pan left, tilt up, truck right, pedestal down, arc, tracking shot, static shot. Add amplitude and speed when it matters, as in the camera pushes in slowly with small amplitude, but do not decorate every move. A static shot is often the strongest choice, provided you write it down rather than leaving the field empty.
Sound
Handle three layers separately, because they fail separately. Dialogue: exact words in quotation marks, attached to a named speaker, inside your word budget. Scene sound: physical noises tied to visible action, such as cutlery, footsteps, rain on glass. Background music: describe it or ban it outright. Declaring that non-diegetic music is not applicable is the difference between clean room tone and a stock score. This is the block a MiniMax prompt guide written before native audio does not have.
The exclude list
Three to six specific negatives, not a wall of generic quality words. No soft dissolves, no lens flare, no added on-screen text, no misspellings, no extra people in frame. Specific exclusions work; a pile of adjectives does not.
Pick the mode before you write
Mode selection is a control decision, and the cheapest lesson in any MiniMax H3 prompt guide: getting it right removes work from the prompt. A MiniMax H3 prompting guide that ignores this makes you do the same work in prose.
Text to video gives the model everything to invent, so your prompt carries the whole visual load. Choose it for concepts with no fixed identity, and set the aspect ratio explicitly: text mode will not infer one.
First and last frame mode pins the opening or closing composition to an image you already trust. The model fills the motion between and will not invent cuts, which is the cheapest way to stop endings drifting. Both frames must share an aspect ratio, and the output inherits it.
Omni-reference mode is the full twelve-file workflow, and it beats the other two only if every file has a declared role. Uploaded without roles, more references make results less predictable, not more.
Output lands at 2K by default at twenty-four frames per second, with a 768P option and ratios from 21:9 through 9:16. Test low, finish high.
Three MiniMax H3 video prompt examples
Each MiniMax H3 video prompt example below is compressed for reading. Expand the timeline block when you use them.
The first MiniMax H3 video prompt example is text only, six seconds:
Shot on 35mm, fine grain, warm tungsten light, quiet domestic mood. [0-3s] An older man in a wool cardigan lifts a kettle from the stove. [3-6s] He pours, steam rises across the frame, he watches it. Camera holds a static wide shot. Sound: kettle hiss, water pouring, faint refrigerator hum. Non-diegetic music: none. Exclude: on-screen text, lens flare, additional people.
The second pins both ends, eight seconds:
Image 1 is the exact opening frame. Image 2 is the exact closing frame. Preserve the subject's jacket, glasses, and hair. [0-4s] She turns from the window toward the desk. [4-8s] She sits and settles into the closing pose. Camera pushes in slowly with small amplitude. Sound: chair movement, distant traffic, controlled breathing. Non-diegetic music: not applicable. Exclude: cuts, dissolves, added props.
The third MiniMax H3 video prompt example is worth studying, because role assignment carries the detail. Ten seconds, omni-reference with dialogue:
Image 1: speaker identity only. Image 2: workshop location and colour grade only. Video 1: hand motion and timing only. Audio 1: vocal timbre only, not the words. Style: handheld documentary, cool daylight, shallow depth of field. [0-4s] He tightens a clamp on a workbench, glances off-camera. [4-10s] He looks to the lens and says, "This is the part everyone gets wrong." Camera: subtle handheld, one slow push in. Sound: metal on wood, workshop air, no music. Exclude: added dialogue, on-screen captions, second speaker.
Twenty-two words of dialogue would have been the ceiling at ten seconds. That line is nine, which leaves room for the glance, the pause, and the push in.
Diagnose the clip, not the prompt
When output disappoints, the useful question is which block failed, not whether the prompt was bad. This table is the part of the MiniMax H3 prompt guide you will return to.
- The character drifts between shots. Reference roles. Name three or four stable anchors and check that identity images are not competing.
- The opening frame is not your image. Mode. A loose reference is not a pinned frame; switch to first-frame mode.
- Music appears that you never asked for. Sound. Add an explicit ban rather than leaving the layer unmentioned.
- The wrong character speaks. Sound. Map each voice to one named speaker at its first appearance and keep the identifier consistent.
- Everything feels rushed. Budget. Cut one action or one cut; do not shorten what remains.
- On-screen text is garbled. Exclude list. Quote the exact string and forbid misspellings and invented characters.
- Motion ignores your reference video. Reference roles. A busy clip offers too many candidate motions; supply one readable action.
- The ending drifts. Mode. Pin a last frame.
Work that list in order before rewriting from scratch, which is what most MiniMax H3 prompts get instead. A MiniMax H3 prompt guide earns its place by shortening the loop, not by producing longer drafts.
A test ladder that protects your credits
Escalate deliberately. Run pass one at four to six seconds with a static camera, no dialogue, and only the identity reference: you are confirming subject and setting, nothing else. Add one motion instruction in pass two, voice and scene sound in pass three, a second character or a cut in pass four. Only then generate at full duration and resolution.
Each pass isolates one variable, so when something breaks you know which addition broke it. This is the least glamorous part of any MiniMax H3 prompt guide and the one that saves the most money: slower on paper, far faster than regenerating a fifteen-second clip six times while changing four things at once.
Where this MiniMax H3 prompt guide stops applying
This MiniMax H3 prompting guide assumes you want controlled, repeatable output: a specific character, a specific beat, a specific sound. If you are exploring visually and want to be surprised, over-specification is counterproductive, and a MiniMax prompt guide aimed at open-ended work would tell you the opposite: short, evocative, one strong reference.
Block structure matters less as reference material gets stronger. A character sheet carrying shot order and framing already encodes what the timeline block would say, and repeating it in prose only creates room for contradiction. Let the strongest artefact carry the detail.
Endpoint limits differ by platform, particularly maximum duration, so confirm what your host allows before planning a fifteen-second sequence. The principles hold; the ceilings move.
Put the budget to work
Prompting here is a scheduling skill more than a writing skill. Decide what fits in the seconds you have, give every reference one job, say something deliberate about sound. Those three lines are the MiniMax H3 prompt guide worth memorising, and they beat a longer prompt that ignores any of them. That is a complete MiniMax H3 prompting guide in three sentences.
Judgment comes from iterating with intent, so run the ladder on a real idea. On SoraLum you can take this minimax h3 prompt guide straight into text-to-video, start with a six-second test pass, and work up to the full clip without switching tools. Bring one action and one line of dialogue, and a few passes will tell you whether your budget was honest.
