What is the best MiniMax H3 prompt structure?

Write reference roles first, then the visible scene and shot timeline, then exclusions and preservation rules. For multi-shot video, give each shot a time range, composition, action, camera instruction, dialogue, and sound.

How should I reference uploaded files in an H3 prompt?

Assign one job to every file, such as identity, object, location, style, composition, motion, camera, voice, music, or a strict first or last frame. Refer to it explicitly as Image 1, Video 1, or Audio 1.

How do I ask for no background music?

State the constraint explicitly at the end of the prompt, for example: non_diegetic_music: N/A. Do not also request a music mood elsewhere in the prompt.

Can H3 preserve exact text and logos?

Exact text is fragile. Supply important typography as a reference image, quote the required wording in the prompt, keep motion simple, and plan to composite legal or brand-critical text in post.

How to write MiniMax H3 prompts

The practical formula

Use three layers: reference instructions, what should be visible, and what must not appear. For the visible layer, specify shot size, subject, environment, action, camera, dialogue, synchronized sound, and music. Write concrete images rather than metaphors.

Assign every reference one role

Label each upload explicitly: Image 1 owns character identity, Image 2 owns the setting, Video 1 owns action rhythm, and Audio 1 owns voice or emotional cadence. If one feature must remain exact, name it. Conflicting owners create drift.

Write a timeline for every cut

Use chronological time ranges. A reliable shot contains one composition, one main action, and at most one camera instruction. Make dialogue short enough for its shot. You can request a J-cut or L-cut when a line or sound crosses an edit.

Use physical camera language

Name the move, direction, amplitude, speed, and timing. Instead of “orbit,” write truck left plus pan right, or truck right plus pan left. State whether the sequence uses normal cuts, dissolves, rhythm-matched cuts, fast cuts, aerial movement, slow motion, or one continuous take.

Quote dialogue and on-screen text

Write the exact spoken line and identify the speaker. Quote any title, logo, button label, or screen text that must appear. For critical typography, use an image reference and tell H3 to interpret it as an image rather than regenerating the characters.

Remove contradictions

Do not request a one-take sequence and then describe several cuts. Do not ask for music and later forbid background music. Do not upload a file without declaring its purpose. When no reference exists, give H3 enough visual material: subject appearance, setting detail, action, camera, sound, and ending.

Complete timed example

0-3 seconds: wide shot. The woman from Image 1 walks slowly into the cherry-blossom courtyard from Image 2. Shallow background focus. No dialogue, footsteps only. 3-8 seconds: cut to a medium shot of the same woman. She draws the sword from Image 3 and begins a measured form while petals fall. Slow push in. 8-12 seconds: cut to a close-up. The blade flashes in slow motion and scatters the petals. non_diegetic_music: N/A.