How to prompt MiniMax H3 for text-to-video: action chains, sound cues and resolution choices from 768p to 4K. A template, three example prompts and mistakes.
MiniMax H3 works from a prompt alone or from a first and last frame, and it always adds sound. So it pays to write prompts where the action and the sound belong together: something breaks, metal rings, water moves. This guide covers how MiniMax H3 reads a prompt, a template built around cause and effect, three text-to-video examples and the mistakes to avoid.
With a prompt alone, MiniMax H3 works as text-to-video. Add a first frame to fix the opening image, and a last frame to fix where the shot ends. MiniMax H3 References takes up to 9 reference images instead of frames. They must be photographic, and line art is refused.
Clips run from 5 to 15 seconds at 768p, 2K or 4K. Sound is always on, with no switch to turn it off, so every prompt is also a sound brief, whether you write one or not.
MiniMax H3 Max is the sibling lane. It is faster, follows the prompt more tightly, renders 480p to 1080p, offers true text-to-video, and accepts frames and up to 9 references together. Sound is always on there too.
[Shot size and speed]: [subject] [first action], [what that causes], [the result]. [Setting and light]. Sound: [the key effect], [the ambience].
Lead with a short label that sets the shot, such as "Close-up in slow motion:" or "Sharp aerial shot". Then write the action as a chain, with one physical beat leading to the next. End with the sound, because MiniMax H3 produces audio either way.
Made in Cinemagiq. The exact prompt is under each one.
PromptClose-up in slow motion: a glass of red wine is knocked off a table, the wine arcs through the air and the glass shatters on a stone floor.
PromptA blacksmith hammers a glowing horseshoe on an anvil in a dark forge, sparks fly with every strike, embers float in the air. Sound: heavy metallic clangs and a roaring fire.
PromptSharp aerial shot over a turquoise glacier lagoon with floating icebergs, a tiny red kayak crossing the still water, crisp morning light.
Example 1 opens with a two-part label, "Close-up in slow motion:", that sets framing and speed before the action starts. The action is a cause-and-effect chain: the glass "is knocked off a table", "the wine arcs through the air" and "the glass shatters on a stone floor". There is no Sound line, so this prompt leaves the soundtrack to the model, which always adds one.
Example 2 ties motion to rhythm. "Sparks fly with every strike" links the visual effect to a repeated action, and "embers float in the air" adds slow motion against the fast hammering. The Sound line, "heavy metallic clangs and a roaring fire", names one rhythmic effect and one constant bed of sound.
Example 3 was set to 2K, and its words match the setting: "Sharp aerial shot" and "crisp morning light" both ask for detail. A wide landscape gets a single focal point, "a tiny red kayak crossing the still water", so the eye has somewhere to go.
For more on shot sizes, camera moves and sound cues, see the prompt engineering cheat sheet.
Yes. Sound is always on for MiniMax H3 and MiniMax H3 Max, so every clip comes with audio and there is no switch to turn it off. Write a Sound line naming the effects and ambience you want. If you leave it out, the model chooses the audio itself.
MiniMax H3 renders at 768p, 2K or 4K, for clips of 5 to 15 seconds. Use 768p for drafts and blocking, then 2K or 4K for the take you keep. MiniMax H3 Max covers 480p to 1080p instead.
Yes. MiniMax H3 References takes up to 9 images, but only photographic references, and it refuses line art. MiniMax H3 Max can combine first and last frames with up to 9 references in one generation, which helps when a framed shot also needs a consistent character.
MiniMax H3 Max is faster, follows prompts more tightly and renders 480p to 1080p, with true text-to-video and frames plus up to 9 references together. MiniMax H3 offers 768p, 2K and 4K. Both run from 5 to 15 seconds with sound always on.
Run MiniMax H3 and the other leading models in one AI-native workspace.