MiniMax H3 Prompt Guide: Structure, Examples and Real Clips

An H3 prompt has three fields: integrated_multimodal_description for the visuals and timed shots, overall_soundscape for ambience, and non_diegetic_music for music only viewers hear. Put At 00:05.000, the camera cuts to … before each later shot, tag speakers as (S1) and (S2), and wrap each line in <d>[English] …</d>. We ran seven prompts on October 28, 2026, and every cut we asked for landed within a quarter second.

The MiniMax H3 prompt structure

MiniMax's model card says H3-Base consumes the structured block, and calls its rewriting step, H3-Context-IR, "critical to the quality of the final output". MiniMax never makes the structure mandatory, and its own examples use plain sentences too.

• Opening line (image-to-video and first-and-last-frame only): one sentence naming the frames and their time marks, then one blank line; the first line of the prompt.

• integrated_multimodal_description: shots with timecodes, actions, camera, speakers and dialogue; the first of three core fields.

• overall_soundscape: ambient, action and non-verbal sounds, 1–4 English sentences; N/A only when the user asks for complete silence.

• non_diegetic_music: music only the audience hears, 1–3 English sentences; N/A when there is no such music.

• For image-to-video, the first line is the fixed sentence that names the frame, followed by one blank line.

• Each later shot starts with a timecode, such as At 00:05.000, the camera cuts to ….

• The first speaker is (S1), and each line sits inside <d>[English] …</d>.

• Put visible signs and labels in English double quotation marks.

• Give overall_soundscape 1–4 English sentences and non_diegetic_music 1–3. Write N/A only when you want complete silence, and name one style label in shot 1, such as "Live-action, cinematic".

How we made the clips

• We ran minimax-h3 at 768P through OmniaKey on October 28, 2026. Text-to-video clips are 16:9; image-to-video clips take their shape from the first frame.
• The first frames came from GPT Image 2 at 1536×1024, medium quality. The last frame of the tulip clip is an edit of its first frame.
• Each clip took 30 to 50 seconds to render. When we sent five jobs at once, one was rate-limited; sent alone, it rendered.
• We read the frames at fixed times, measured where each cut fell, how long each file ran and how loud it was, and had two audio models transcribe the speech.
• OmniaKey sells API access to MiniMax H3, so we ran the clips through it. The prices below are MiniMax's list prices, not ours.
• Sources: MiniMax's base and ref prompt guides, checked October 28, 2026, and its pricing page, checked October 26, 2026.

We ran one clip per prompt, so treat each result as one sample.

Seven prompts with real clips

1. A plain sentence and the structured version of the same bakery

The same bakery scene, written two ways, shows what the three fields add.

Plain sentence:

A baker opens the shutters of his small street bakery before sunrise, puts a fresh loaf on the wooden counter and says "First batch of the morning." Then a close-up of steam rising from the sliced bread.

Structured fields:

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium-wide shot frames a baker opening the shutters of a small street bakery before sunrise. The camera pushes in with small amplitude at slow speed as the middle-aged baker with a calm, slightly raspy voice (S1) places a fresh loaf on the wooden counter and says: <d>[English] First batch of the morning.</d> [Shot 2] At 00:05.000, the camera cuts to a close-up of steam rising from the sliced bread while the baker's final words carry over from the previous shot.

overall_soundscape: Wooden shutters scrape open over a quiet street as trays clink softly inside the bakery. The doorbell rings once, followed by light footsteps and the crisp sound of bread being sliced.

non_diegetic_music: A soft acoustic-guitar pattern at a moderate tempo, joined by sparse upright-bass notes and a gentle fade at the end.

Settings: text_to_video, 768P, 16:9, 8 seconds each.

Result: the structured version cut on time; the plain version did not, and its sound is quieter.

• The structured clip cut at 5.00 seconds, the time its prompt asks for. The plain clip cut at 6.54 seconds, where no time was asked.
• The plain prompt has no sound fields. Its track averages −43.4 dB, against −18.7 dB for the structured clip.
• Two audio models disagreed on the plain clip, and one heard words the prompt does not contain, so we do not quote a spoken line for it.

What to change:

• Try a timecode for every cut you need, even in a plain prompt.
• Try putting any sign text you want read in English double quotation marks. Both bakery clips had unquoted signs, and those came out as letters, not words.

2. Three timed shots, cuts at 3.5 and 7 seconds

Clip two of the set is the multi-shot test: three shots in a 10-second clip, with each cut written as a timecode.

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a small fishing harbor at dawn, where an old fisherman in a yellow raincoat coils a wet rope on a wooden pier. The camera pushes in with small amplitude at slow speed. [Shot 2] At 00:03.500, the camera cuts to a close-up of his weathered hands tying the rope to an iron cleat. [Shot 3] At 00:07.000, the shot cuts to a low-angle shot as three seagulls lift off the pier railing and the fisherman looks up, shading his eyes with one hand.

overall_soundscape: Small waves lap against the wooden posts while the rope creaks as it tightens. Seagulls cry overhead and their wings flap as they take off.

non_diegetic_music: N/A

Settings: text_to_video, 768P, 16:9, 10 seconds; non_diegetic_music set to N/A.

Result: the first cut landed on time, the second came 0.25 seconds late, and the file ran 10.13 seconds.

• The cut at 3.50 seconds matches the prompt. The cut at 7.25 seconds does not match the 7.00 seconds it asks for.
• The cleat shot is on screen by 5.0 seconds, and the gulls lift off at about 8.5 seconds.
• The track averages −39.4 dB, and neither audio model heard speech.

What to change:

• Try ending shot 2 a quarter second earlier if the second cut must land on 7.00 seconds.

3. Two speakers in one scene

Clip four of the set has two speakers, each with one line. Kling 3.0 writes dialogue differently, so for that format see our Kling 3.0 prompt guide.

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a medium shot frames a tiny ramen counter at night under a red neon sign reading "OPEN UNTIL 2 AM". The cook, a middle-aged man with a deep, warm voice (S1), sets a steaming bowl in front of a young woman in a green jacket and says: <d>[English] One more bowl?</d> [Shot 2] At 00:04.000, the camera cuts to a close-up of the young woman with a bright, quick voice (S2), who lifts her chopsticks and says: <d>[English] Make it extra spicy.</d>

overall_soundscape: Broth simmers in a large pot and a ladle clinks against the rim. Light rain patters on the awning outside.

non_diegetic_music: N/A

Settings: text_to_video, 768P, 16:9, 8 seconds; non_diegetic_music set to N/A.

Result: the sign read correctly and both lines came through, but the cut came 0.08 seconds early.

• Both audio models transcribed "One more bowl. Make it extra spicy." word for word.
• The cook serves the bowl in shot 1, and the cut to the young woman comes at 3.92 seconds.
• The sign reads OPEN UNTIL 2 AM in both shot-1 frames we checked.

What to change:

• Try quoting every sign you want read, not just one. The one quoted sign in this clip came out right.

4. Image-to-video: the first frame sets the look

Clip five starts from a still we made with GPT Image 2 at 1536×1024: a woman at a pottery wheel in a studio with window light. Our GPT Image 2 prompt guide is the place to start for stills.

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 woman shown in <Picture 1> stays seated at the pottery wheel, preserving her short black hair, denim apron, the grey clay and the studio shelves behind her. The wheel starts to spin and she presses her thumbs into the clay, opening it and drawing the walls upward into a tall cylinder. The camera pushes in with small amplitude at slow speed toward her hands. She looks up and smiles, and the woman with a calm, low voice (S1) says: <d>[English] Almost there.</d>

overall_soundscape: The wheel motor hums steadily and wet clay squelches under her fingers. Water drips softly into the splash pan.

non_diegetic_music: N/A

Settings: image_to_video, 768P, 6 seconds requested, first frame, no aspect ratio.

Result: the woman kept her look from the first frame, and the clay rose into a taller form.

• Her hair, apron and the studio shelves match the still in every frame we checked.
• The clay is a cone at 2.0 seconds and a tall form with a flared top at 4.0 seconds. At 6.4 seconds she is smiling.
• Four audio models heard "Almost there."
• The file runs 6.58 seconds, not 6.

What to change:

• Try planning a 6-second image-to-video cut as about 6.6 seconds.

5. First and last frame: a tulip opens

Clip seven has a first frame and a last frame, both still images. The first is a closed red tulip bud in a glass vase; the last is an edit of it, with the flower open.

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, a static shot begins with the closed red tulip bud in the glass vase established by Picture 1. In soft morning light the outer petals loosen and curl outward one by one, the bud widens, and the inner petals part to reveal the yellow center, until the flower settles into the fully open shape, composition and lighting established by Picture 2 at the end of the shot.

overall_soundscape: Quiet room ambience with faint birdsong outside the window.

non_diegetic_music: Sparse piano notes at a slow tempo.

Settings: image_to_video with first_frame and last_frame, 768P, 6 seconds requested.

Result: the bud opened to the last frame, with the camera still and the vase unchanged.

• The petals are part-open at 3.0 seconds and fully open at 6.5 seconds, matching the last frame.
• Both audio models heard solo piano, and neither heard speech.
• The file runs 6.58 seconds, and the track averages −26.0 dB.

What to change:

• Try a longer clip for a slower opening. We ran only one 6-second version.

6. Silence: both sound fields set to N/A

Clip six tests what N/A does: both sound fields set to N/A, with a fox walking through snow.

integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a red fox walking slowly across fresh snow in a pine forest at first light. The camera tracks the fox at slow speed as it stops, lifts its head and looks toward the camera.

overall_soundscape: N/A

non_diegetic_music: N/A

Settings: text_to_video, 768P, 16:9, 6 seconds requested; both sound fields N/A.

Result: the clip is close to silent, but not silent.

• The track averages −57.8 dB and peaks at −38.2 dB.
• At that level the audio models' transcripts are not reliable, so we report no sounds for this clip.
• The file runs 6.58 seconds.

What to change:

• Try N/A in overall_soundscape only when you want complete silence, as the guide says.

Common failures and fixes

These come from our own clips, not from the guide. The section numbers refer to the prompts above.

What a clip costs

MiniMax bills per output second at its list prices.

A 10-second clip costs $0.80 at 768P and $1.30 at 2K, calculated from the same rates. Extra input images cost $0.04 each, and the first 5 per task are free. H3-Context-IR, the optional rewriting call, is billed by tokens; its example response reports 9,090 total tokens.

MiniMax list prices, checked October 26, 2026. OmniaKey's rates are on the MiniMax H3 model page, and the call flow is in our MiniMax H3 API guide.

Run these prompts through the API

This request sends the harbor prompt from section 2. Blank lines in the prompt become \n inside the JSON.

curl https://api.omniakey.com/v1/media/tasks \
  -H "Authorization: Bearer $OMNIAKEY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: harbor-shots-001" \
  --data-binary @- <<'JSON'
{
  "model": "minimax-h3",
  "type": "video",
  "operation": "text_to_video",
  "input": {"prompt": "integrated_multimodal_description: [Shot 1] Live-action, cinematic, a wide shot frames a small fishing harbor at dawn, where an old fisherman in a yellow raincoat coils a wet rope on a wooden pier. The camera pushes in with small amplitude at slow speed. [Shot 2] At 00:03.500, the camera cuts to a close-up of his weathered hands tying the rope to an iron cleat. [Shot 3] At 00:07.000, the shot cuts to a low-angle shot as three seagulls lift off the pier railing and the fisherman looks up, shading his eyes with one hand.\n\noverall_soundscape: Small waves lap against the wooden posts while the rope creaks as it tightens. Seagulls cry overhead and their wings flap as they take off.\n\nnon_diegetic_music: N/A"},
  "parameters": {"resolution": "768P", "duration": 10, "aspect_ratio": "16:9"}
}
JSON

1. The reply is 202 with a task id.
2. Poll https://api.omniakey.com/v1/media/tasks/{id} with that id until status is final. Our clips took 30 to 50 seconds. Sending the same key and body again returns the same task.
3. Download the file from assets[0].url. That link is signed and lasts about an hour, so poll again for a fresh one if it expires.

For image-to-video, set operation to image_to_video and put the still in input.media with type first_frame, as an HTTPS URL or a data URI. The whole request body is capped at 1 MiB, including inline Base64. The model page and the video API docs cover the rest.

Common questions

Does MiniMax H3 need the structured format?

No. MiniMax never makes the structure mandatory, and H3-Context-IR, its paid rewriting step, turns plain input into the structured block for you. Our plain-sentence clip still produced a video, but its cut was not timed.

How long can an H3 prompt be?

The limit is 7,000 characters per text item. The structured bakery prompt above is far below that.

Can I write H3 prompts in Chinese?

MiniMax lists 11 languages with stable dialogue support, Chinese among them, but it publishes no list of languages for the prompt text itself. The guide asks for the six rewrite sections in English and keeps dialogue in its own language, and our prompts were in English.

Can I turn the sound off in H3?

MiniMax documents no mute setting. Putting N/A in both sound fields gave a near-silent clip in our test, and the guide allows N/A in overall_soundscape only when you want complete silence.

Why did my sign come out as gibberish?

Signs the prompt did not quote came out as unreadable letters in our clips. The guide asks for visible sign text in English double quotation marks, and the one quoted sign we ran read correctly.

Update log

• October 28, 2026: first version, seven prompts run the same day.