MiniMax H3 prompt guide

This guide follows MiniMax’s prompt writing guide for H3 and its API guide. For the model, see the MiniMax H3 page.

MiniMax’s guide splits a prompt into three parts: a description along the timeline (style, composition, subjects, action, camera, speakers and the sounds tied to an action), a soundscape for the whole clip, and music the characters can’t hear.

1. 1Open with the style and first composition: “Live-action, cinematic, a medium-wide shot frames…”.
2. 2Name each subject with the details that must stay: age, clothes, colors, position.
3. 3Write the action in order, with the camera move inside the sentence.
4. 4Start each later shot with its cut time: “[Shot 2] At 00:05.000, the camera cuts to…”.
5. 5Describe each voice once, then the line in quotes, with its language.
6. 6Close with the soundscape, then the music, or say there is none.

MiniMax’s guide writes the parts as labeled fields, the form its prompt rewriter, H3-Context-IR, produces; its API examples are plain sentences. Either way, keep the order.

MiniMax’s guide builds a move from its type, how far it goes (amplitude) and how fast (speed), written as an action in the sentence, not a label at the end.

MiniMax’s API guide also accepts a short form: [pan], [zoom] or [static] straight after the description it applies to. Hailuo’s camera guide adds: one main move per clip, calmed with slow or subtle.

For a steady frame, Hailuo’s guide to unwanted motion suggests rig words, a locked-off shot or a tripod; handheld adds randomness. If only the distance changes, MiniMax says move the camera rather than cut.

Every H3 clip has stereo sound made with the picture. MiniMax puts lines and action sounds in the description, where they happen; the soundscape covers the rest; music goes last. See native audio.

• Describe a voice once: age, pitch, timbre, pace or accent. MiniMax gives each speaker an ID, (S1) or (S2), kept across shots.
• Keep the delivery outside the quotes and only the spoken words inside: she says in French, “Bonjour, les croissants sortent du four.”
• For a voice-over, write “says in an off-screen voiceover”, then add that the on-screen character’s lips stay closed.
• Write the soundscape in one to four sentences: wind, footsteps, breathing; no dialogue.
• Write the music as instruments, tempo and volume, not moods, or say there is none.

MiniMax’s model card lists dialogue that holds up in 11 languages: Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish; others work less consistently. Put any sign or label text in double quotes, exactly as it should read. See lip sync.

MiniMax’s API takes up to 7,000 characters, and its rewriter’s prompts run to hundreds of words: spend them on what is seen and heard, moment by moment.

Time cuts as MiniMax does: no time on the first shot, a later time on each one after, and something new at each cut. Hailuo’s general prompt guide suggests 20 to 50 words for most platforms; H3’s own examples run longer, so try both.

On DreamBench, H3 renders 21:9, 16:9, 9:16 for Reels and Shorts, 1:1 and 4:3; MiniMax’s Hailuo AI app also offers 3:4.

MiniMax’s API takes a first frame, a last frame or both; the picture sets the clip’s shape. See image to video and first and last frame.

1. 1Start from the picture: the young woman by the rainy train window.
2. 2Say how the action begins: the camera trucks right slowly as she looks up from the letter.
3. 3End on a result: she folds the letter along its crease.

MiniMax’s own example asks for change: “Pull focus to the people in the background and add more steam to the ramen bowl.” DreamBench’s composer has no start or end slot: in a film it draws each opening frame from your references, and the prompt says how the clip ends.

MiniMax’s reference mode takes up to 9 images, 3 videos and 3 audio clips, 12 files in all; its feature examples name each in the prompt: “Image 2: character. Image 3: bag.” References and first or last frames can’t be mixed.

• Describe each character in the same words in every shot: “a middle-aged baker with a gray beard and a green apron”.
• Keep each speaker’s voice description, and their ID, the same across shots.
• Keep the clothes, props, time of day and light.

DreamBench doesn’t send H3 reference images, videos or audio. In a film, each character, place and prop is drawn once as a sheet, and each clip’s opening frame is drawn from the sheets it needs. See character consistency and reference to video.

MiniMax trained H3 for multi-shot video, and its guide times cuts inside one clip. Past 15 seconds, or to redo one shot alone, make each shot its own clip. See multi-shot video.

• Inside one clip: number the shots, time each cut after the first, describe subjects identically.
• Across clips: one moment each, new framing at every join, ambience running across it.

On DreamBench, describe the film and pick MiniMax H3: it writes each shot as one continuous clip of up to 12 seconds, draws its opening frame and cuts the clips together; redo any one with a note. On the Video page, your shot labels reach H3 as written.

Paste these into DreamBench’s Video page or any MiniMax H3 tool. Each is one clip.

Three timed shots: the morning bakery

12 seconds, 16:9. [Shot 1] Live-action, cinematic, a medium-wide shot of a street bakery before sunrise. A middle-aged baker with a gray beard and a green apron sets a loaf on the counter as the camera pushes in slowly. The baker, a calm, raspy voice, says in English: “First batch of the morning.” [Shot 2] At 00:05.000, the camera cuts to a static close-up of steam rising as he slices the loaf. [Shot 3] At 00:09.000, the shot cuts to a customer’s hand taking a paper bag. Soundscape: trays clink, the doorbell rings once. Music: a soft acoustic guitar, fading at the end.

Product with a label: a tin of tea

8 seconds, 1:1. Live-action, a close shot of a tin of tea on an oak table, morning light from the left. The lid reads “HIGHLAND BREAKFAST” in gold capitals. A hand lifts the lid and a spoon scoops dark leaves; the camera pushes in slowly, the lettering sharp. Soundscape: the lid’s soft pop, leaves rustling. Music: none.

Dialogue: two speakers on a night bus

10 seconds, 16:9. Live-action, a static two-shot on the top deck of a night bus, city lights streaking past. A young man in a yellow raincoat, a quick, nervous voice, says in English: “I think we missed our stop.” The woman beside him, a low, amused voice, doesn’t look up from her book: “Two stops ago.” Soundscape: the engine’s hum, rain on the roof. Music: none.

A line in Spanish: the fruit stall

6 seconds, 9:16. Live-action, a static medium shot at a market stall in midday sun. An older woman in a striped apron holds a mango up to the camera and, in a warm, brisk voice, says in Spanish: “¡Prueba este, está dulcísimo!” Soundscape: market chatter, a paper bag crackling. Music: none.

Voice-over: a letter read aloud

10 seconds, 16:9. Live-action, a close shot of an old man in a brown cardigan reading a handwritten letter by lamplight. The camera trucks left slowly. The old man, a soft, low voice, says in an off-screen voiceover: “I still remember the road to your house.” His lips stay closed. Soundscape: a clock ticking, rain at the window. Music: sparse piano at a slow tempo.

Close-up sound: bread on a board

6 seconds, 9:16. Live-action, a static extreme close-up of a sourdough loaf on a board. A serrated knife saws through it in three slow strokes and the slice falls away. Soundscape: the crust crackling under the blade, close and crisp. Music: none.

Stereo travel: a tram across the frame

10 seconds, 21:9. Live-action, a wide static shot of a cobbled square at dusk. A red tram enters on the left, rattles across and leaves on the right, ringing its bell once. Soundscape: the rumble travels from left to right in stereo, wheels squeal on the bend. Music: none.

Animation: a clay robot

8 seconds, 1:1. Claymation, a small blue clay robot with one round eye on a workbench. It picks up a bolt, screws it into its own chest and gives a proud bounce. An arc shot circles it slowly. Soundscape: clay squeaks, a metallic click. Music: a light xylophone at a bouncy tempo.

App motion: a banking screen

8 seconds, 9:16. 3D CG, a phone screen on a lilac background. A balance card slides up reading “£2,480.00”, three payments drop in, and a green check mark appears beside “Rent paid”. The camera pushes in slowly. Soundscape: soft taps, one chime on the check mark. Music: a minimal electronic pulse.

Sport: a rower at dawn

8 seconds, 16:9. Live-action, a sculler in a white tank top rows across a misty lake at dawn. A tracking shot follows at water level, then the camera pedestals up as the boat glides past. Soundscape: oars dipping, the seat sliding, a bird calling. Music: none.

Vertical trailer: the lighthouse

15 seconds, 9:16. [Shot 1] Live-action, cinematic, a wide shot of a lighthouse as storm waves break over it. [Shot 2] At 00:04.000, the camera cuts to a young keeper with cropped red hair and an oilskin coat at the rail; she shouts in English, in a hoarse voice: “The light’s gone out!” [Shot 3] At 00:09.000, the shot cuts to her hand striking a match in the dark lamp room. Soundscape: wind howling, waves booming. Music: low strings building, stopping on the flame.

A one-line prompt with a camera instruction

5 seconds, 16:9. A fox trots across a snowy field at dusk [static], stops, then dives head first into the snow [zoom]. Soft wind, a muffled thump.

Most weak H3 prompts are short, not wrong.

• One prompt run twice gives two clips.
• MiniMax says visual detail can still improve in some scenes: keep key details large in frame, at 1080p.
• Every clip has sound; for silence, ask for room tone and mute it in an editor.
• A real person or product can’t be known from words; it takes a picture you have the right to use.
• One generation stops at 15 seconds, and on DreamBench’s Video page a prompt past 5,000 characters is refused.

• One clip: on the Video page, pick MiniMax H3, paste your prompt, choose 5, 6, 8, 10, 12 or 15 seconds, the frame and 720p or 1080p. It goes to H3 as written.
• A prompt: the AI video prompt generator writes one for any model in shot-list order; it has no H3 preset and writes the prompt only.
• A film: describe it, add up to 8 reference pictures and pick MiniMax H3; DreamBench writes, draws and cuts every shot, and you can redo any one.
• Not offered: first or last frames, references sent to H3, video editing, or 3:4.

MiniMax H3 is on every plan. The composer shows the price in credits before you generate; the first time, create a free account and choose a plan. For another model, see the Grok Imagine 1.5, Veo 3.1 or Kling 3.0 prompt guide; for prompts from words alone, text to video.