MiniMax H3

With reference images or video inputs connected, start and end frames are not available. A single prompt therefore uses either the keyframe format or the reference format, never both.

• Sound has no on/off setting and is always generated together with the video, so background sound and music are also set in the prompt's sound fields.
• A prompt can be up to 7,000 characters long. When you supply the prompt by connecting a text node, anything beyond 7,000 characters is not sent to the model either.

You can check the supported inputs, length, ratio, and resolution in Video models.

Write the three fields in the order below, with a blank line between them.

integrated_multimodal_description: [Shot 1] ...

overall_soundscape: ...

non_diegetic_music: ...

Write all three fields in English. The only things kept in their original language are the dialogue inside <d> and text visible on screen.

The alignment sentence is fixed for the frames you connect. Leave one blank line below it, then write the three fields to complete the prompt.

When only a start frame is connected

For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced.

When a start frame and an end frame are connected

How the reference pictures align with the target video — Picture 1 (from Shot 1) aligns with the 0.00-second mark of the target video; Picture 2 (from Shot N) aligns with the S.SS-second mark of the target video.

• N is the number of the actual last shot, and S.SS is the video length written to two decimal places. If you set the length to 8 seconds in the video node, it is 8.00.
• The start frame is Picture 1, and the end frame is Picture 2.
• The end frame image belongs to the last shot. Do not attach it to Shot 1.
• When you use a start frame and an end frame together, it is best to build the video as a single shot.
• When you use a prompt alone, do not write an alignment sentence.

The young woman with a quiet, breathy voice (S1) says: <d>[English] I get off at the next station.</d>

For a voice heard from off screen, write exactly says in an off-screen voiceover, and state right after it that the character's mouth is closed. This keeps mouth movement from being wrongly added to a character on screen.

How references are numbered

• Reference images become <Picture 1>, <Picture 2>, and so on, in the order you connected them to the video node. An image that you disconnect and connect again takes the last number.
• For video inputs, videos connected on the canvas are numbered first, followed by videos uploaded directly to the node.

Open [Shot 1] by stating the style and opening composition. Describe the style with words such as Live-action, cinematic, 2D-animated, 3D CG, claymation, watercolor, or vintage film.

From the second shot on, write the cut time, and make each shot's time later than the one before.

[Shot 2] At 00:03.500, the camera cuts to a close-up of her hands.

Cut only when the character, space, state, point of view, or time changes. To change only the distance or angle, move the camera instead of cutting.

Write camera movement as part of a sentence, without bracketed commands.

The camera pushes in with small amplitude at slow speed toward the letter in her hands.

The available movements are Zoom In/Out Push In/Pull Out Pan Left/Right Truck Left/Right Tilt Up/Down Pedestal Up/Down Arc Shot Tracking Shot Static Shot Shake Slightly/Strongly POV Roll Clockwise/Counterclockwise.

Add amplitude (with small/large amplitude) and speed (at slow/fast speed) only when needed. They do not have to go on every movement.

The body section is named detailed_description, not integrated_multimodal_description.

Choose a label

Use <Subject> when you draw on something inside a reference, and <Picture> or <Video> when the file itself fills a role.

Example

Connect a photo of a person and then a photo of a café as reference images, and make an 8-second video in two shots of a woman sitting alone in the café and answering a phone call. The photo of the person was connected first, so it is <Picture 1>, and the café photo is <Picture 2>.

Step 1 — subject_definitions

Write one line for each thing to track separately, not for each file. The two photos define a person and a place rather than becoming specific frames, so do not give <Picture N> its own entry; cite it only as the source inside the Subject definitions.

subject_definitions:
<Subject 1> is the young woman in <Picture 1>, with long dark hair and a blue cardigan.
<Subject 2> is the café interior in <Picture 2>, with a brick wall and a wooden counter.

Step 2 — summary

Start with the task type in square brackets, then describe what you are making in one paragraph. Do not introduce new labels here.

Connecting a file does not set the type automatically. Even with a video connected, if you reference only its camera movement, the type is reference generation.

This example builds a person and a place from photos, so the type is reference generation.

summary:
[reference generation] The target video shows <Subject 1> sitting alone in <Subject 2> and answering a phone call across two shots.

Step 3 — retention_analysis

Write one line for each label defined in step 1, in the order which shots it appears in → marker → what is kept.

retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - her identity, long dark hair, and blue cardigan are retained.
<Subject 2> (appears in [Shot 1]): partially_preserved - the brick wall and wooden counter are retained, but the lighting shifts to evening.

• Choose markers only within the role defined in step 1. The café lighting shifting to evening changes an attribute of the place itself, so it is partially_preserved. The woman answering the phone adds an event that was not there before rather than changing her attributes, so she stays fully_preserved.
• Do not use speaker numbers such as (S1) in this section.

Step 4 — detailed_description

This is the body of the video. Describe the visuals, action, camera, dialogue, and sound in the scene in playback order.

• Start with one or two sentences on style, then begin [Shot 1] on the next line.
• Place labels where they first appear and wherever their role applies.
• Do not add a time to [Shot 1]. From the second shot on, write the cut time in the At 00:04.000 format.

detailed_description:
The target video is in a cinematic, muted-color style with soft evening light.
[Shot 1] A medium-wide shot establishes <Subject 2>, the café interior with its brick wall and wooden counter, now lit by low evening light from the street. <Subject 1> (S1), the young woman with long dark hair and a blue cardigan, sits alone at the counter with a phone face-down beside her cup. The camera pushes in with small amplitude at slow speed as the phone vibrates and she turns it over.
[Shot 2] At 00:04.000, the shot cuts to a close-up of <Subject 1> (S1) raising the phone to her ear. She answers, <d>[English] I didn't think you'd call.</d> She closes her lips and looks toward the window.

• When a character speaks, attach both the label and the speaker number (<Subject 1> (S1)). The label identifies who the character is, and the speaker number identifies who is speaking.
• Describe an important Subject's features only when it first appears, and in later shots write just the label. Do not define it again.
• A length of 350–500 English words works well. The example above is shortened to show the structure; in practice, describe each shot's composition, lighting, action, and sound in more detail.

Step 5 — the two sound sections

In overall_soundscape, describe the background sound, sounds of movement, and nonverbal sounds of the whole video in 1–4 English sentences. In non_diegetic_music, describe the background music that only the audience hears in 1–3 sentences, focusing on instruments, tempo, and rhythm. If there is no music, write N/A.

overall_soundscape: Low café room tone continues throughout, with a distant espresso machine and quiet street traffic outside. The phone vibrates once against the wooden counter.

non_diegetic_music: Sparse piano notes at a slow tempo, joined by a sustained low string that fades out before the final line.

Write dialogue and lyrics only inside <d> in the body, and do not repeat them in the sound sections.

Complete prompt

This is the prompt with all five steps combined.

subject_definitions:
<Subject 1> is the young woman in <Picture 1>, with long dark hair and a blue cardigan.
<Subject 2> is the café interior in <Picture 2>, with a brick wall and a wooden counter.

summary:
[reference generation] The target video shows <Subject 1> sitting alone in <Subject 2> and answering a phone call across two shots.

retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2]): fully_preserved - her identity, long dark hair, and blue cardigan are retained.
<Subject 2> (appears in [Shot 1]): partially_preserved - the brick wall and wooden counter are retained, but the lighting shifts to evening.

detailed_description:
The target video is in a cinematic, muted-color style with soft evening light.
[Shot 1] A medium-wide shot establishes <Subject 2>, the café interior with its brick wall and wooden counter, now lit by low evening light from the street. <Subject 1> (S1), the young woman with long dark hair and a blue cardigan, sits alone at the counter with a phone face-down beside her cup. The camera pushes in with small amplitude at slow speed as the phone vibrates and she turns it over.
[Shot 2] At 00:04.000, the shot cuts to a close-up of <Subject 1> (S1) raising the phone to her ear. She answers, <d>[English] I didn't think you'd call.</d> She closes her lips and looks toward the window.

overall_soundscape:
Low café room tone continues throughout, with a distant espresso machine and quiet street traffic outside. The phone vibrates once against the wooden counter.

non_diegetic_music:
Sparse piano notes at a slow tempo, joined by a sustained low string that fades out before the final line.