MiniMax H3 Max
Text to Video

Sample output

Made With MiniMax H3 Max — Prompt Included

Press any card and its prompt drops into the box above, with the ratio and length already set to match. Every prompt on this wall is the one its publisher printed beside the clip — fal for the English ones, metaso's MiniMax H3 case gallery for the rest, translated from Chinese here and otherwise unchanged.

• Level 1

  6 words

  A woman walks down a street.

  Drifts

  The subject changes between frames. The camera picks its own move. You did not say, so it decided.

  Run this one

• Level 2

  30 words

  00:00:12:0030 words

  A woman in a red coat walks down a wet Tokyo alley at night, neon reflected in puddles, slow push-in with small amplitude, rain on metal and distant traffic.

  Usable

  The camera does what you asked and the rain is actually in the audio. This is the level most good prompts sit at.

  Run this one
• The structure the model actually reads

  Level 3

  Three fields

  00:00:12:00Official fields

  integrated_multimodal_description

  [Shot 1] Red coat, wet Tokyo alley, slow push-in, small amplitude.

  overall_soundscape

  Rain on metal, distant traffic, footsteps in standing water.

  non_diegetic_music

  Low synth pad, slow pulse, sitting under the dialogue.

  Deliverable

  The spoken line sits in front of the music because you put it there. Nothing is left for the model to guess.

  Open this in the builder

Rather not write three fields? The MiniMax H3 Max prompt generator builds all of them from one sentence.

• Balanced

  Default

  About one second

  Cleans up your sentence and fills the obvious gaps. Keeps the whole point of this model, which is that you get the clip back before you have finished re-reading the prompt.

• Quality

  Up to 30 seconds, before rendering starts

  Writes a much richer prompt. On a 5-second clip that rewrite is longer than everything else put together — worth it on a final take, wasteful on the nine drafts before it.

Both figures are the model provider's own, from its published API schema.

Then it shows you the rewrite. Every clip you make here can expand to reveal the prompt the model actually received — the full rewritten text, word for word.
Most sites hand you the video and keep the rewrite. Reading it once teaches more than any guide will.
You see what MiniMax H3 Max filled in for you, and which parts you should have pinned down yourself.
Copy it, edit it, run it again as your own Level 3 prompt.

The rewrite panel under a finished clip, with all three fields expanded

1. 01Workflow

   Paste the Whole Script

   The box takes 7,000 characters, not one line. Press Paste a script above and drop the lot in.

2. 02Workflow

   Cut It Into Shots

   Number them `[Shot 1]`, `[Shot 2]`, and timestamp the cuts. Pick any whole length from 5 to 15 seconds for each one.

3. 03Workflow

   Render Each Shot

   Six aspect ratios, every whole second, one shot at a time. Draft at 480P and re-run the keepers at 768P.

4. 04Workflow

   Chain Them Together

   Take the last frame of one clip and feed it in as the first frame of the next in image to video. That is how a character survives a cut.

   What we do not do: timeline editing. Four 12-second clips make 48 seconds, and the joining happens in your CapCut or Premiere. Those already do it better than we would, and we would rather say so than sell you a worse one.

• 21:9Cinematic

  Title cards, banners, widescreen previz

  Try this ratio
• 16:9Widescreen

  YouTube, landing pages, the default

  Try this ratio
• 4:3Classic

  Retro and archive looks

  Try this ratio
• 1:1Square

  Product cards, marketplaces, email

  Try this ratio
• 3:4Portrait

  Portrait feeds and Pinterest

  Try this ratio
• 9:16Vertical

  TikTok, Reels, Shorts

  Try this ratio

There is no `adaptive` for text to video. Image to video can read the frame off the picture you upload; text to video has no picture to read. Leave the ratio blank and you get 16:9.

Ratio does not change what you pay. A 9:16 clip and a 21:9 clip of the same length cost the same.

• 5sEnough for one action
• 10sEnough for a beat and a reaction
• 15sThe longest single take

• ·Draft at 480P, finish at 768P. A 5-second draft is 125 credits; the same shot as a keeper is 200. That is the intended way to work here, not a downgrade.
• ·Resolution is the only multiplier. Nothing else moves the rate — not the ratio, not the audio, not the length of your prompt.
• ·A run that fails costs nothing, on every plan including the free one.

00:00:10:00Two speakers · 768P · 10s

The rule most people miss

Keep the descriptive half of your prompt in English, but write each spoken line in its own language.

• (S2) replies firmly, [Japanese] 分かった。Right
• (S2) replies firmly, [Japanese] Understood.Wrong

Writing a Japanese line out in English is the single most common reason a delivery comes back wrong.

Need the same face across several clips? Anchor it with a still in image to video.

Source: Official · MiniMax API reference

Four things that trip people up

• 01

  Quality expansion eats the whole speed advantage

  Balanced rewrites your prompt in about a second. Quality spends up to thirty before rendering even starts.
  On a model whose entire point is that it comes back fast, that is the difference between a draft loop and a coffee break. Leave it on Balanced and switch only for a final take.

• 02

  There is no adaptive ratio here

  Image to video reads the frame off the picture you upload. Text to video has nothing to read it from, so it falls back to 16:9. Pick the ratio deliberately — it is free to change and it changes the whole composition.

• 03

  The safety check is on, and some ordinary shots get caught

  Real people by name, graphic injury and some brand marks come back rejected. A rejected run does not cost you credits. If a normal creative shot gets caught, rewording the subject usually clears it — see what we allow.

• 04

  The specs you find elsewhere disagree with each other

  We checked. Third-party wrappers and resale docs are quoting their own defaults as if they were the model's.

  	The model's own schema	Seen on third-party docs
  Shortest clip	5 seconds	3 seconds
  Aspect ratios	6 options	7, including one that does not exist
  Default resolution	768P	480p
  Default length	5 seconds	8 seconds

  Defaults are a wrapper's choice, not a model property — ours are 480P and 5 seconds, because drafts should be cheap. The first two rows are simply wrong, and a request built on them comes back as an error.

1. 1

   Write the picture, then the sound

   One field takes both. Describe what the camera sees, then what it hears — dialogue, effects, score, in that order.

   Up to 7,000 characters
2. 2

   Set the ratio, the length and the size

   Six ratios, any whole second from 5 to 15, and 480P or 768P. The button shows what the run costs before you press it.

   24 fps · 5–15 seconds
3. 3

   Download the MP4

   One file with the sound already in it. A run that fails costs no credits, on every plan including free.

   No watermark, on any plan

That is the whole browser path: nothing to install, no GPU, no API key.

• Alternatives

  Best AI video generator with sound: which ones charge you extra for it

  Nearly every serious video model generates audio now. What still separates them is whether sound costs extra and how many takes your budget buys. Eight, priced.

  Editorial deskSep 4, 2026
• Prompt craft

  Six ad variants by Friday: how I write MiniMax H3 Max prompts now

  The MiniMax H3 prompt guides are for a model H3 Max does not include. The five-part brief I use instead, six controlled variants, and the frame grid.

  Editorial deskSep 4, 2026
• Alternatives

  Kling or MiniMax H3 Max: the queue is the part nobody prices

  Kling reaches native 4K and multi-shot storyboards. It also costs three to ten times more per second, and the waiting is a real line item.

  Editorial deskSep 4, 2026

• Dialogue prompts
• Product prompts
• Cinematic prompts
• Browse all prompts