Most text-to-video prompts fail the same way: they describe a picture, not a shot. You get a pretty frame that barely moves, or motion that drifts into mush. MiniMax H3 Max rewards the opposite habit — write like a director. Tell it where the camera starts, where it ends, how the seconds are spent, and what the scene sounds like, and it holds the shot together with real camera logic and a matching audio track.
This guide is three copy-paste recipes, each built around one control you actually need: camera movement, pacing over a fixed duration, and choreographing several moving elements at once. Every clip below was generated on hiapi with the exact prompt printed next to it — nothing cherry-picked, nothing cleaned up. Copy a prompt, change the subject, ship your own.
All three samples were rendered at 480P for fast, cheap iteration (the cheapest tier); the same prompts scale straight up to 768P and 1080P when you want delivery quality.
How MiniMax H3 Max reads a prompt
The call takes a prompt plus three knobs, and the prompt itself carries a fourth, invisible one — sound. Get the mental model right and the recipes below stop feeling like magic:
prompt— a shot description, not a caption. The model responds to camera verbs ("push in", "crane up", "orbit"), to lighting and mood words, and to an explicitAmbient sound:line that drives the generated audio.duration— a number from0.92to15seconds. This is your pacing budget. More seconds means you can stage beats; fewer means commit to one move.resolution—480P,768P, or1080P. Iterate cheap, deliver sharp.aspect_ratio—21:9,16:9,4:3,1:1,3:4, or9:16. Pick for the destination (cinematic, social vertical, square product tile) before you write the prompt, because framing changes how you describe the composition.
Two habits matter across every recipe. First, anchor the camera: say where it begins and where it ends, and the motion in between stays coherent. Second, list your negatives — no people, no text, no cuts, no speech — because text-to-video models love to hallucinate captions and hard cuts into a shot that should be one continuous take.
Recipe 1 — Camera movement: the push-in and crane reveal
The single most useful move is starting tight and opening up. Here the camera begins inches from a red umbrella on wet stone, then pushes forward and cranes up to reveal the whole alley. One subject, one continuous move, a payoff at the end.
Single continuous shot. A lone open red umbrella lies on wet cobblestones
in an empty European alley at dawn, thin fog, puddles mirroring amber
streetlamps. The camera starts low and close on the umbrella, then slowly
pushes forward and cranes upward into a wide reveal of the long alley
stretching toward a misty cathedral. Gentle natural motion: drifting fog,
a few slow raindrops rippling the puddles. Cinematic, shallow depth of
field, warm-cool color contrast, stable geometry, seamless temporal
consistency, no people, no text, no cuts. Ambient sound: soft rain,
distant church bells, faint wind, no speech.
Parameters: aspect_ratio: 16:9 · duration: 6 · resolution: 480P
Why it works:
- Two camera positions, one sentence. "starts low and close… then pushes forward and cranes upward into a wide reveal" gives the model a start frame and an end frame. It interpolates a believable move between them instead of wandering.
- Motion that isn't the camera. "drifting fog" and "raindrops rippling the puddles" give the scene life without competing with the camera move. Secondary motion keeps the frame from looking frozen.
- Mood words do real work. "warm-cool color contrast" and "shallow depth of field" push the look toward cinematic; the amber-vs-fog palette in the result came straight from that line.
- Swap test: replace the umbrella and alley with "a vintage motorcycle in a neon-lit garage" and keep the camera sentence verbatim — the move transfers cleanly.
Recipe 2 — Pacing: spending a fixed duration with timed beats
When you have more than a few seconds, don't describe one static moment for the whole clip — script the beats. Labeling time windows inside the prompt tells the model how to spend the duration. This 8-second vertical runs a three-beat pour sequence.
Vertical social clip, one continuous take, three beats.
Beat 1 (0-3s): a barista's hands pour hot espresso over a glass of ice,
close macro, dark crema swirling.
Beat 2 (3-6s): milk cascades in slow motion, blooming into feathered
clouds through the coffee.
Beat 3 (6-8s): the camera pulls back to reveal the finished iced latte on
a sunlit marble counter beside fresh mint.
Warm morning light, crisp condensation on the glass, smooth natural
motion, high detail, stable focus, no text. Ambient sound: gentle pour,
ice clink, soft cafe hum, no speech.
Parameters: aspect_ratio: 9:16 · duration: 8 · resolution: 480P
Why it works:
- Beats map to seconds. "Beat 1 (0-3s)… Beat 2 (3-6s)… Beat 3 (6-8s)" tells the model to allocate its duration, so the action progresses instead of looping one gesture. Keep your beat windows adding up to the
durationyou request. - One beat = one change. Each beat introduces exactly one new thing (pour → milk bloom → pull-back reveal). Stack two changes into one beat and the model tends to rush or blur them.
- Reserve the reveal for last. The pull-back in Beat 3 is the payoff. Ending on the wider frame reads as a deliberate finish, not a clip that ran out of time.
- Vertical is a composition choice, not just a crop. With
9:16you describe a tall frame — a glass filling the center — rather than a landscape squeezed into a phone screen.
Recipe 3 — Scene choreography: multiple elements moving together
The hard case is several things moving at once without the shot falling apart. The trick is to give each element its own verb and its own path, then give the camera one simple job. Here: a silk ribbon spirals one bottle, mist drifts across the back, and the camera does a slow orbit.
Square product shot, single continuous take. On a dark reflective stage,
three frosted glass perfume bottles stand in a row. A ribbon of golden
silk floats in from the left and spirals around the center bottle, while
a soft burst of white mist drifts across the background and settles. The
camera orbits slowly clockwise around the bottles. Studio rim lighting,
deep shadows, physically accurate reflections on the glass and stage,
premium luxury commercial finish, controlled elegant motion, stable
geometry, no people, no logos, no text. Ambient sound: low cinematic hum,
soft whoosh of silk, delicate chime, no speech.
Parameters: aspect_ratio: 1:1 · duration: 6 · resolution: 480P
Why it works:
- Each element gets a verb and a direction. The silk "floats in from the left and spirals around the center bottle"; the mist "drifts across the background and settles." Naming the path per element keeps them from colliding into noise.
- Give the camera the easy job. With the subjects doing the choreography, a plain "orbits slowly clockwise" is enough. Asking the camera to do something complex and animating three elements is where shots break.
- Physics words buy you realism. "physically accurate reflections on the glass and stage" is why the bottles mirror correctly on the floor — reflective materials need to be called out explicitly.
no logosearns its place. On product shots the model will otherwise invent branding. Negatives are how you keep a clean, usable plate.
Run these yourself
Every clip above is one asynchronous task on hiapi's /v1/tasks endpoint — submit, poll, download. The input fields are exactly the four knobs from the recipes:
curl -X POST https://api.hiapi.ai/v1/tasks \
-H "Authorization: Bearer $HIAPI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "minimax-h3-max/text-to-video",
"input": {
"prompt": "Single continuous shot. A lone open red umbrella on wet cobblestones... Ambient sound: soft rain, distant church bells, no speech.",
"duration": 6,
"resolution": "480P",
"aspect_ratio": "16:9"
}
}'
The response returns a taskId; poll GET /v1/tasks/{taskId} until status is success, then grab the clip from the output URL (it's a short-lived link — download it right away).
Cost scales with the seconds you request, not the seconds you get back. As of October 2026, MiniMax H3 Max text-to-video is billed per requested video second: $0.072/s at 480P, $0.115/s at 768P, and $0.229/s at 1080P. So the three 480P clips here (6s + 8s + 6s) cost about $1.44 total. Iterate your prompt at 480P, then re-run the keeper at 1080P. Live numbers are on the hiapi pricing page.
Already have the first frame you want? The same prompt grammar works on MiniMax H3 Max image-to-video — hand it a still and describe the motion and sound instead of the whole scene.
FAQ
Does MiniMax H3 Max generate sound, or do I add it later?
It generates audio from your prompt. The Ambient sound: line in each recipe is what produced the rain, the cafe hum, and the silk whoosh in the clips above — no post-production. Say no speech unless you specifically want dialogue, and describe the ambience you want rather than leaving it to chance.
What's the maximum clip length? Duration runs from 0.92 to 15 seconds per task. For anything longer, generate several clips and stitch them — and script each one's beats as in Recipe 2 so they cut together cleanly.
Why do my clips look soft or washed out? Usually it's the resolution tier, not the prompt. These samples are 480P for cheap iteration. Re-run your final prompt at 768P or 1080P for delivery sharpness. Also add material and lighting words ("physically accurate reflections", "studio rim lighting") — they sharpen the look as much as the pixel count.
How do I stop it from adding text or captions?
Put no text (and no subtitles, no logos for product work) in your negatives. Text-to-video models frequently hallucinate on-screen text; listing it explicitly is the reliable fix.
Which aspect ratio should I use?
Match the destination: 16:9 or 21:9 for cinematic and landscape, 9:16 or 3:4 for social and mobile, 1:1 for product tiles and feeds. Choose it before writing the prompt, since framing changes how you describe the composition.
Start directing
The pattern under all three recipes is the same: name the camera's start and end, spend the duration in beats, give every moving element its own verb, and write the sound. Steal these prompts, swap in your subject, and iterate at 480P until the move is right.
When it is, re-run it at 1080P and pull the final clip from the MiniMax H3 Max text-to-video model page — same prompt, delivery quality.
