MiniMax H3 Max prompt ideas, printed in full

Every prompt complete — no truncation, no paraphrase, no tidying up. Free to copy, no account. 7 templates

Long briefs, beat by beat

Reference render — fal.ai
P—01H3 MaxOne shot
Collapsing sea-cathedral
A film-school shot brief written as a prompt: named subjects, timed beats, one continuous take.
15s16:9768P$1.20
Reference render — fal.ai
P—02H3 MaxOne shot
Pixel-art endless runner
A game design document used as a video prompt. The rules are what make it read as gameplay.
15s16:9768P$1.20

One shot, written properly

Reference render — fal.ai
P—03H3 MaxOne shot
Night-market dolly
One camera move, one street, one sound list. The shortest prompt here that still directs everything.
5s16:9768P$0.40
Reference render — fal.ai
P—04H3 MaxOne shot
One line to camera
A single line to camera. The words are quoted exactly, which is the entire trick.
5s16:9768P$0.40

Stop motion, slow motion, macro

Reference render — fal.ai
P—05H3 MaxOne shot
Claymation frog with pancakes
Three physical details and a frame rate turn a description into stop motion.
5s16:9768P$0.40
Reference render — fal.ai
P—06H3 MaxOne shot
Hummingbird, slow motion
Slow motion requested deliberately, on the one kind of subject that needs it.
5s16:9768P$0.40
Reference render — fal.ai
P—07H3 MaxOne shot
Espresso, macro
Commercial macro with a subject that transforms on camera, and a four-source mix.
5s16:9768P$0.40

Every prompt here comes from fal's own showcase for MiniMax H3 Max — fal post-trained the model, so these are the prompts the people who tuned it reached for to demonstrate it. The clips are theirs and each card says so; the prompts are what this page publishes. Six of the seven have fal's render beside them. The seventh is fal's API example, which it publishes as code rather than as a clip.
Duration, ratio and cost on a card are the settings the prompt is written for and what a run costs here, not a claim about how fal ran it. The shot brief is the one to watch: it schedules fifteen seconds inside its own text, and the clip fal published against it runs five.

One request

All seven are one prompt. None of them need a picture first.

Every entry here is text to video on MiniMax H3 Max, which is the shortest path there is from an idea to a clip with sound in it. It is worth seeing what the same seven clips would cost to assemble on a model that only makes pictures.

Anywhere else — three passes

A still, then a silent clip, then a soundtrack

Any image modelAny silent video modelA sound pass

A silent video model needs a first frame to hold the look, so there is an image pass in front of it. It returns picture without audio, so there is a sound pass behind it — foley, ambience, score, and a sync edit to land them on the cuts. Three tools, three bills, and two places for the look to drift between them.

Every dialogue clip needs a fourth pass for lip sync.

Here — one pass

Straight to the clip, sound included

MiniMax H3 Max — picture and sound in one pass

H3 Max writes picture and audio together, so a sentence about a wet street and a wok sizzling comes back as both, already in sync. That is why six of these prompts spend their last line on sound: it is an input here, not a separate project. A 5-second clip at 768P is ready in 8.7 seconds at the median, measured on this site.

Copy the text, set duration and ratio, run it.

Want one that is yours?

These seven are a starting point, not a ceiling. The MiniMax H3 Max prompt generator builds the same structure from one line of your own — free, no account, and it checks the result before you copy it.

Generate free · queue

Before you copy anything

MiniMax H3 Max prompt ideas FAQ

Where these came from, and what to check before you paste one in

Where do these MiniMax H3 Max prompts come from?

From fal's own showcase for MiniMax H3 Max. fal post-trained the model on top of the open-weight MiniMax H3 base, and these are the prompts it published to demonstrate what it tuned for — six in the showcase gallery and one in its API example. We republish the text verbatim and link each card back to the source. What we add is the reading: which parts of each prompt are load-bearing, what you can swap, and what stops working if you remove it.

Were these clips generated on this site?

No, and every card says so. The clips are fal's renders from its showcase, kept beside the prompt so you can see what the text is capable of before spending anything on it — six of the seven have one, and the seventh is fal's API example, which it publishes as code with no clip attached. The prompt is what this page publishes. When we run one here ourselves, the label on that card changes.

Will I get exactly the same video?

Not unless you fix the seed. Running the same prompt twice gives you two different takes — same subject, same shot, same sound design, different execution. The H3 Max endpoint does accept a seed, so pinning one is how you get a repeatable starting point. Treat the clip beside a prompt as what that text is capable of rather than what you are guaranteed, and when a take is close, change one thing and run it again rather than rewriting the whole prompt.

What can I change without breaking one?

Nouns are safe: the subject, the place, the colour, the time of day. Structure is not. Shot numbers, cut timestamps and quoted dialogue are read structurally, and removing them changes what the model thinks you asked for. The other common way to break one is adding an action without adding seconds — that usually comes back in slow motion, because the model fits the extra beat in by slowing everything down. Every card lists its own safe edits and its own breaking ones, with the reason attached.

Do all of these run on MiniMax H3 Max?

Yes. All seven are text to video, which is one of the two modes H3 Max supports, and the engine badge on each card confirms it. Nothing here needs reference input — keeping a specific face, product or style by uploading it — because that is not on the H3 Max API. Those jobs run on MiniMax H3 instead, at the same rate per second and with 2K available on top.

Why does every prompt end with a line about sound?

Because audio is generated in the same pass as the picture, so describing it is an input rather than a post-production step. Leave it out and the model still generates sound — it just invents it. Naming the sources gets you a mix instead of a guess.

How long can a prompt be?

7,000 characters. The longest here is just under 5,000, and longer is generally better with this model. Full specs.

Do I need an account to copy these?

No. Copying is free and anonymous. You only need an account when you actually render something. What free covers.

How do I write one of my own?

Three fields in a fixed order: the picture, then what the room sounds like, then the score. The prompt generator assembles that from one line and checks it before you copy.

Why seven MiniMax H3 Max prompt ideas and not two hundred?

Because seven is how many the model's own post-trainer publishes complete. A prompt you cannot read in full is not an example, it is a screenshot, and a library padded to a round number with prompts nobody can trace is a longer page and a worse one. Every entry here is printed whole — the longest runs past 4,900 characters — and carries its source link. The library grows when a prompt earns its place.

Will these work on other video models?

Partly. The prose travels; shot timestamps and field names usually do not, and a model without native audio will ignore every sound line. What survives.

Take any of them. They are yours wherever you paste them.

Seven MiniMax H3 Max prompts, printed in full and free to copy with no account. Press the button on a card and it loads into the tool that runs it, ratio and length already set.