The Formula
Instruction line, blank line, then integrated_multimodal_description + overall_soundscape + non_diegetic_music.
- Instruction line
- For image-to-video, always: For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.
- integrated_multimodal_description
- Everything visible and audible on the timeline: style, composition, subjects, actions, camera motion, cuts, dialogue and on-screen sound. Opens with [Shot 1] and the overall style.
- overall_soundscape
- 1 to 4 sentences of ambience, physical action sounds and non-verbal human sounds. No dialogue here.
- non_diegetic_music
- 1 to 3 sentences of score the characters cannot hear: instrumentation, tempo, rhythm, dynamics. N/A when there is none.
The Rules
Start with the exact instruction line
For image-to-video the first line is fixed wording that pins your image at 0.00 seconds. Follow it with one blank line before the fields.
Open Shot 1 with the style
Begin the description with the overall style, such as "Live-action, cinematic", then the composition from your first frame, then what happens next.
Write camera moves as type, amplitude and speed
Use a natural sentence: "The camera pushes in with small amplitude at slow speed toward her hands." Leave out amplitude and speed when they are medium or normal.
Keep the speaker outside the dialogue tag
Describe the speaker and give a stable ID like (S1) outside <d>. Inside <d> put only the language tag and the exact words, never translated or rewritten.
Mark voiceover explicitly
Use the phrase "says in an off-screen voiceover" and add, right after the <d> block, that the on-screen character's lips remain completely closed.
Keep dialogue out of the soundscape
overall_soundscape is for rain, footsteps, fabric, breathing and room tone. Speech and singing already live in the description.
Describe music concretely
non_diegetic_music names instruments, tempo and dynamics, not mood words. Write N/A when the shot has no score.
Time any cuts
The first shot has no timestamp. Later shots start with "[Shot 2] At 00:03.500, the camera cuts to..." with strictly increasing times; for a small reframe, use camera motion instead of a cut.
A Template You Can Copy
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. integrated_multimodal_description: [Shot 1] [Style, e.g. Live-action, cinematic], [composition and subject as they appear in the first frame]. [Action that develops forward from the frame]. The camera [move] with [small or large] amplitude at [slow or fast] speed [toward what]. The [speaker description and voice] (S1) says: <d>[Language] [exact words]</d> overall_soundscape: [1 to 4 sentences: ambience, action sounds, non-verbal sounds] non_diegetic_music: [instruments, tempo and dynamics, or N/A]
A Real Example
The shot: Night, rain-streaked diner window. A woman in a red coat sits alone in the last booth, turning a ring on the table. Slow push in from the window. Mara says: "You kept it." (6 seconds, 16:9)
For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. integrated_multimodal_description: [Shot 1] Live-action, cinematic night diner scene. A woman in a red coat sits alone in the last booth, rain streaking down the window beside her, warm interior light glowing against the cold dark outside. She turns a ring slowly between her fingers on the tabletop, eyes lowered. The camera pushes in with small amplitude at slow speed, moving from outside the rain-streaked window toward her face and hands. As it settles close, the young woman with a quiet, steady voice (S1) says: <d>[English] You kept it.</d> overall_soundscape: Rain ticking softly against the glass, a faint hum of refrigeration or a neon sign, the low murmur of an otherwise empty diner, and a soft clink as the ring taps the tabletop. non_diegetic_music: A sparse, melancholic piano motif, slow tempo, soft dynamics, underscoring the quiet moment.
It uses the exact instruction line, opens Shot 1 with the style, and states the push in as type, amplitude and speed. The speaker's identity and voice sit outside <d> with an (S1) ID, the rain and ring sounds stay in overall_soundscape, and the score is described by instrument and tempo (the guide would also drop the mood word "melancholic").
Dialogue
Describe the speaker (age, voice, on or off screen) and give a stable ID such as (S1) outside the tag, then put the language tag and the exact words inside <d>...</d>. Use (S1,S2) for lines spoken together, "says in an off-screen voiceover" for voiceover, and never repeat dialogue in overall_soundscape.
The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>
Camera Moves
| Move | How to write it for MiniMax H3 |
|---|---|
| Slow push in | The camera pushes in with small amplitude at slow speed toward the folded letter in her hands. |
| Fast pan | The camera pans right with large amplitude at fast speed, revealing the open doorway. |
| Static | The camera holds a static shot as the runner exits the frame. |
| Truck | The camera trucks right with small amplitude at slow speed as she lifts her gaze. |
| Pull out | The camera pulls out with small amplitude at slow speed as she opens the umbrella. |
| Arc | The camera moves in an arc shot around him with small amplitude at slow speed. |
Common Mistakes
| Mistake | Fix |
|---|---|
| Skipping the instruction line on an image-to-video prompt. | Start with the fixed line that pins <Picture 1> at 0.00 seconds, then a blank line. |
| Writing the line as plain quoted text. | Put the words in <d>[English] ...</d> and the speaker description with (S1) outside the tag. |
| Repeating dialogue in overall_soundscape. | Keep only ambience, action sounds and non-verbal sounds there. |
| Mood words in non_diegetic_music, such as "sad, emotional score". | Name instruments, tempo and dynamics: "sparse piano notes at a slow tempo, fading out". |
| Stacking camera labels at the end of a sentence. | Write the move as part of the action: "The camera pushes in with small amplitude at slow speed toward her hands." |
Length, Resolution and Price
MiniMax H3 makes 5 to 15 second clips with sound always on and lip-synced dialogue, and it accepts an end frame. It is billed per second at $0.05 at 480p, $0.06 at 768p, $0.13 at 2K (its native resolution) and $0.16 at 4K. Reference mode takes up to 9 images, the first 5 free and $0.08 each after that (checked 2026-10-22). Every model side by side is on the model spec sheet.