Explore tested MiniMax H4 prompts for cinematic videos, product advertisements, character scenes and creative AI workflows. Customize prompts and generate AI videos directly.
MiniMax H4 prompts are detailed text instructions used to guide AI video generation. Effective prompts describe the subject, action, environment, camera movement and visual style to help MiniMax H4 create more accurate and cinematic video results. Strong prompts can also define shot timing, character consistency, reference roles, dialogue and sound. Start with a clear scene, then add only the motion and visual details that support the intended workflow.
PROMPT LIBRARY PLAYGROUND
Generate Videos From MiniMax H4 Prompts
The best way to understand a prompt is to test the result. Customize these MiniMax H4 prompts and generate your own AI videos directly.
Cinematic Drone Shot
Cinematic Video
Selected
Use case: A text-only scene with a clear visual start, a timed second shot, and full soundscape plus background music.
integrated_multimodal_description:
[Shot 1] Live-action cinematic night scene. A woman in a dark raincoat walks alone along a narrow train platform as heavy rain falls beyond the roof. A medium-wide tracking shot moves backward in front of her at walking speed. Reflections from blue platform lights move across the wet ground. Her shoes make small splashes as she walks.
[Shot 2] At 00:04.500, the camera cuts to a close-up from her left side. She stops, looks toward the approaching train and slowly closes her umbrella. White headlights grow brighter behind her as wind moves loose strands of her hair.
overall_soundscape:
Steady rainfall, wet footsteps, distant rail vibration, low station ambience and the rising metallic sound of an approaching train.
non_diegetic_music:
A restrained electronic pulse with sparse low synth notes, gradually increasing in intensity.
Image-to-Video Product
Product Advertisement
Prompt Card
Use case: Start from the first frame, preserve what must stay fixed, and spend the prompt on what changes after it.
For the target video, at 0.00 seconds, <Picture 1> is the opening frame.
integrated_multimodal_description:
[Shot 1] Begin from <Picture 1> and preserve the character's face, clothing, color palette and initial composition. The woman slowly turns toward the window as wind begins moving the curtain behind her. The camera pushes in gently while sunlight shifts across her face. She lifts one hand and touches the glass.
overall_soundscape:
Quiet indoor room tone, soft fabric movement and distant city traffic through the window.
non_diegetic_music:
N/A
First/Last Frame Transformation
Character Consistency
Prompt Card
Use case: One continuous shot that explains the path from the opening frame to the target final frame.
How the reference pictures align with the target video:
Picture 1 is the opening frame and Picture 2 is the final frame.
integrated_multimodal_description:
[Shot 1] Begin from Picture 1. Preserve the woman's identity and clothing as she slowly walks from the center of the empty daylight street toward the camera. The camera pulls backward at the same speed. As she moves, shop lights switch on one by one, the sky darkens, reflections appear on the pavement and the scene gradually transforms into the neon night environment shown in Picture 2. Her movement remains continuous without a cut and the final composition settles precisely into Picture 2.
overall_soundscape:
Light footsteps, distant city traffic gradually becoming busier, soft electrical hum from illuminated signs.
non_diegetic_music:
Minimal electronic ambience slowly builds as the environment changes from day to night.
Dialogue Scene
Dialogue & Character
Prompt Card
Use case: Two speakers with stable speaker IDs and spoken words kept inside the <d> tag.
The young woman with a quiet, breathy voice (S1) says:
<d>[English] We have five minutes before the storm reaches us.</d>
The older man with a low, rough voice (S2) replies:
<d>[English] Then we leave now.</d>
Reference-to-Video
Reference-Based
Prompt Card
Use case: A compact six-section reference prompt that assigns each asset one explicit job.
subject_definitions:
<Subject 1> is the main character whose facial identity and hairstyle come from <Picture 1>.
<Subject 2> is the black jacket and silver accessories shown in <Picture 2>.
<Video 1> provides the walking motion and slow handheld tracking-camera behavior.
<Audio 1> provides the voice timbre for <Subject 1> (S1).
summary:
[reference generation] Generate a new night-street scene that preserves <Subject 1>'s identity and <Subject 2>'s clothing while borrowing the walking motion and camera behavior from <Video 1> and the voice timbre from <Audio 1>.
retention_analysis:
<Subject 1>: preserve facial identity and hairstyle.
<Subject 2>: preserve clothing color, material and accessories.
<Video 1>: reference body motion and camera movement only.
<Audio 1>: reference voice timbre and delivery only.
detailed_description:
The target video uses a cinematic night-street style.
[Shot 1] <Subject 1> wearing <Subject 2> walks through a rain-soaked neon street. The walking rhythm and handheld tracking movement follow <Video 1>. <Subject 1> (S1) looks toward the camera and says: <d>[English] I knew you would find me.</d>
overall_soundscape:
Steady rain, wet footsteps and distant traffic.
non_diegetic_music:
Low electronic ambience with a slow restrained pulse.
Create movie-style scenes with camera movement, lighting and cinematic composition.
Product Advertisement Prompts
Generate commercial videos for products, brands and marketing campaigns.
Character Consistency Prompts
Create stable characters across AI-generated scenes.
Camera Motion Prompts
Control camera movements including tracking shots, zoom and cinematic transitions.
Fantasy And Creative Prompts
Generate imaginative worlds and storytelling scenes.
Social Media Video Prompts
Create short-form videos for TikTok, YouTube Shorts and social platforms.
How To Write Better MiniMax H4 Prompts
Better prompts describe the subject, action, environment, camera movement and visual style clearly.
01
Subject
+
02
Action
+
03
Environment
+
04
Camera Movement
+
05
Visual Style
+
06
Quality Details
Prompt engineering example
Subject: A luxury car
Action: Driving through a futuristic city
Environment: Night city with neon lights
Camera: Slow cinematic tracking shot
Style: Premium commercial film
MiniMax H4 Prompt Tips
Describe Camera Movement
Use camera instructions such as tracking shot, drone view, zoom and cinematic movement.
Add Visual Details
Describe lighting, atmosphere, environment and visual style.
Define Motion
Explain how characters and objects should move.
Improve Consistency
Add detailed descriptions for important characters and objects.
MiniMax H4 Prompt Workflow
01
Choose Prompt
→
02
Customize Prompt
→
03
Generate Video
→
04
Review Result
→
05
Improve Prompt
MiniMax H4 Prompts For Different Users
Creator
Create cinematic videos, social content and creative experiments.
For normal MiniMax H4 T2V, I2V and first/last-frame generation, structure the prompt around three fields: integrated_multimodal_description, overall_soundscape, and non_diegetic_music. The first field describes the video shot by shot, including subjects, actions, camera movement, dialogue and synchronized sound. The second summarizes ambient and physical sounds. The third describes background music that only the audience hears. Reference-to-video uses a more detailed six-section format because each image, video, audio file and reusable subject needs an explicit role.
The biggest practical difference from a generic AI video prompt is that H3 benefits from a timeline. Instead of writing “cinematic woman walking in rain,” specify what appears in Shot 1, how the camera moves, what the character does, what is heard, when a cut occurs, and what changes in the next shot.
MiniMax H4 Prompt Modes at a Glance
MiniMax H4 prompt modes
Mode
What You Provide
What the Prompt Must Do
T2V / T2VA
Text only
Define the complete visual and audio timeline
I2V / I2VA
First image
Start from Picture 1 and describe what happens next
FL2V / FL2VA
First + last image
Describe how the opening frame continuously becomes the final frame
L2V / L2VA
Last image
Build a plausible earlier state and converge toward the supplied final frame
R2V / Ref2VA
Images, video and/or audio references
Assign explicit identities, motion, camera, style and audio roles to references
MiniMax's documentation uses the suffix “A” because H3 generates audio together with video. Users searching for T2V, I2V or R2V are generally referring to the same video workflow families.
Based on MiniMax's current H3 prompt specification.
The Native H3 Prompt Format
For the base H3 generation modes, the prompt body is not simply “subject + action + camera + lighting.” MiniMax's current prompt guide organizes the result into three explicit fields.
This is the timeline of the video. Put the visual style, composition, subject appearance, actions, reactions, shot changes, camera movement, spoken dialogue and synchronized diegetic sound here.
overall_soundscape
Use this field for the sound environment that persists across the video: rain, traffic, wind, footsteps, cloth movement, impacts, breathing, room tone or other physical and ambient sounds.
Do not duplicate spoken dialogue here if it has already been placed inside the relevant shot.
non_diegetic_music
Use this field only for score or background music that the audience hears but the characters inside the scene do not.
Background music example
non_diegetic_music:
Slow sparse piano with sustained low strings and no vocals.
No background music
non_diegetic_music:
N/A
The MiniMaxH4.org Four-Layer Prompt Method
MiniMaxH4.org practical framework
A useful way to build an H3 prompt without losing important details is to write it in four layers.
The four prompt layers
Layer
Question to Answer
Example
1. Scene
What exists at the start?
A woman stands alone on a rain-soaked station platform
2. Action
What visibly changes?
She closes her umbrella and walks toward the arriving train
3. Camera
How does the viewer observe the change?
Slow tracking shot moving backward in front of her
4. Audio
What should be heard at the same time?
Rain, footsteps, distant train brakes and restrained ambient score
Write the four layers as planning notes first, then merge them into H3's native shot timeline. This avoids a common failure mode where a prompt contains many visual adjectives but no executable sequence of events.
How to Write Shots and Timing in MiniMax H4
The first shot begins directly with [Shot 1] and does not need a timestamp. Add timestamps when later shots begin.
[Shot 1] ...
[Shot 2] At 00:03.500, the camera cuts to ...
[Shot 3] At 00:07.200, the shot transitions to ...
Every later timestamp must increase and must fall inside the requested video duration.
When Should You Add Another Shot?
Add a new shot when the viewer receives genuinely new information: a new viewpoint, subject state, location, moment in time or composition. If you only want to move closer, move sideways or reveal something within the same space, camera movement is usually clearer than an unnecessary cut.
Shot Budget by Video Length
MiniMaxH4.org practical starting point
Suggested shot budget by video length
Duration
Suggested Starting Structure
4–5 seconds
Usually 1 clear shot
6–8 seconds
1–2 shots
9–12 seconds
2–3 shots
13–15 seconds
2–4 shots
This is not a hard MiniMax model limit. It is a practical starting point for keeping short H3 generations readable instead of forcing too many scene changes into a few seconds.
MiniMax H4 Camera Movement Guide
Describe camera movement as part of the current shot rather than stacking disconnected camera keywords at the end of the prompt.
Camera movement terminology
Desired Result
Camera Instruction
Move physically closer
Push In
Move physically farther away
Pull Out
Change focal length closer
Zoom In
Change focal length wider
Zoom Out
Rotate view horizontally
Pan Left / Pan Right
Move camera horizontally
Truck Left / Truck Right
Rotate view vertically
Tilt Up / Tilt Down
Raise/lower entire camera
Pedestal Up / Pedestal Down
Circle around subject
Arc Shot
Follow moving subject
Tracking Shot
Keep view fixed
Static Shot
Simulate subject viewpoint
POV
Add physical instability
Shake Slightly / Shake Strongly
Rotate horizon
Roll Clockwise / Roll Counterclockwise
Pan vs Truck, Zoom vs Push: The Difference
Pan and truck are not the same camera movement. A pan rotates the camera horizontally from a fixed position; a truck moves the entire camera sideways. Zoom and push-in are also different: zoom changes focal length while the camera remains in place, while push-in physically moves the camera toward the subject.
Using the correct movement term gives the model a clearer physical instruction than treating all four as interchangeable “cinematic camera movement.”
Control Camera Amplitude and Speed
When the movement needs tighter control, add an amplitude and/or speed modifier.
The camera pushes in with small amplitude at slow speed.
The camera pans right with large amplitude at fast speed.
The camera holds a static shot.
Do not mechanically add amplitude and speed to every camera instruction. Use them only when the range or pacing matters.
How to Write Dialogue for MiniMax H4
Give every speaking character a stable speaker ID such as (S1) or (S2), and reuse that ID across shots.
The young woman with a quiet, breathy voice (S1) says:
<d>[English] We have five minutes before the storm reaches us.</d>
The older man with a low, rough voice (S2) replies:
<d>[English] Then we leave now.</d>
Keep the speaker description and delivery outside the <d> tag. Put only the language tag and the words being spoken inside <d>.
How do I keep the same H3 voice across multiple shots?
Assign the character one speaker ID, such as (S1), when the voice first appears and reuse the same ID every time that character speaks in later shots. Do not renumber the speaker after a cut.
How to Write Voice-over
The man (S1) says in an off-screen voiceover:
<d>[English] I still remember that road.</d>
while his lips remain completely closed.
Explicitly separating an off-screen voice from visible lip movement reduces ambiguity between narration and on-camera speech.
How to Prompt Visible Text
When a sign, label, subtitle, screen or neon board must contain specific text, place the visible text inside English double quotation marks and preserve the exact spelling.
A blue neon sign reading "OPEN ALL NIGHT" flickers above the entrance.
Do not write “the sign says something about being open” if the text must be exact.
A useful rule is to ask whether the characters inside the scene can hear the sound. If they can, it belongs to the scene. If only the audience hears it as a score, it belongs in non_diegetic_music.
Based on MiniMax's current H3 prompt specification.
MiniMax H4 Text-to-Video Prompt Template
T2V has no reference frame, so the prompt must establish the initial visual state as well as the motion and sound.
integrated_multimodal_description:
[Shot 1] Live-action cinematic night scene. A woman in a dark raincoat walks alone along a narrow train platform as heavy rain falls beyond the roof. A medium-wide tracking shot moves backward in front of her at walking speed. Reflections from blue platform lights move across the wet ground. Her shoes make small splashes as she walks.
[Shot 2] At 00:04.500, the camera cuts to a close-up from her left side. She stops, looks toward the approaching train and slowly closes her umbrella. White headlights grow brighter behind her as wind moves loose strands of her hair.
overall_soundscape:
Steady rainfall, wet footsteps, distant rail vibration, low station ambience and the rising metallic sound of an approaching train.
non_diegetic_music:
A restrained electronic pulse with sparse low synth notes, gradually increasing in intensity.
Starter template created by MiniMaxH4.org using the current H3 prompt structure. Do not label it “tested” until a real output is attached to this page.
In I2V, the first image already defines much of the appearance, composition and starting state. The prompt should spend more of its information budget describing what changes after the first frame.
First-frame anchor → action begins → continuous development → visible result
Practical summary based on the official H3 I2VA structure.
For the target video, at 0.00 seconds, <Picture 1> is the opening frame.
integrated_multimodal_description:
[Shot 1] Begin from <Picture 1> and preserve the character's face, clothing, color palette and initial composition. The woman slowly turns toward the window as wind begins moving the curtain behind her. The camera pushes in gently while sunlight shifts across her face. She lifts one hand and touches the glass.
overall_soundscape:
Quiet indoor room tone, soft fabric movement and distant city traffic through the window.
non_diegetic_music:
N/A
This example was written by MiniMaxH4.org using the H3 prompt structure. It is not quoted from MiniMax's official documentation.
What Should You Not Repeat in an I2V Prompt?
Do not spend most of the prompt re-describing information that is already obvious in the first frame unless that detail must remain fixed. Use the text budget to describe motion, state change, camera behavior and sound.
Image says what exists. Prompt says what happens next.
With first-and-last-frame generation, Picture 1 defines the opening state and Picture 2 defines the target ending state. The prompt's job is to explain the path between them.
Think transition, not description.
What to describe in a first-and-last-frame prompt
Describe
Example
Body transition
standing → kneeling
Object transition
closed box → open box
Camera transition
wide → medium close-up
Light transition
daylight → neon night
Scene transition
empty room → crowded room
Pose transition
facing camera → turning away
For this mode, a continuous single-shot transition is often the clearest starting point. Add cuts only when the creative concept genuinely requires them.
MiniMax H4 First-and-Last-Frame Prompt Template
Starter Template
How the reference pictures align with the target video:
Picture 1 is the opening frame and Picture 2 is the final frame.
integrated_multimodal_description:
[Shot 1] Begin from Picture 1. Preserve the woman's identity and clothing as she slowly walks from the center of the empty daylight street toward the camera. The camera pulls backward at the same speed. As she moves, shop lights switch on one by one, the sky darkens, reflections appear on the pavement and the scene gradually transforms into the neon night environment shown in Picture 2. Her movement remains continuous without a cut and the final composition settles precisely into Picture 2.
overall_soundscape:
Light footsteps, distant city traffic gradually becoming busier, soft electrical hum from illuminated signs.
non_diegetic_music:
Minimal electronic ambience slowly builds as the environment changes from day to night.
Full-reference prompting is structurally different from normal T2V or I2V. The current MiniMax H4 reference guide separates the prompt into six sections so that the model can track what each reference means and how much of it should be preserved or transferred.
Six sections of a reference prompt
Section
Purpose
subject_definitions
Defines reusable referenced content and its source
summary
States the target task and main reference relationships
retention_analysis
Explains what is preserved, transferred, copied or only loosely referenced
detailed_description
Describes the target video shot by shot
overall_soundscape
Describes ambience and physical sound
non_diegetic_music
Describes audience-only music
Based on MiniMax's current H3 prompt specification.
Subject, Picture, Video and Audio Labels Explained
Reference label meaning
Label
Use It For
<Subject N>
A reusable person, object, environment, clothing, style, action or visual trait extracted from references
<Picture N>
A concrete reference image used as a frame, keyframe, storyboard or composition anchor
<Video N>
A source video used for motion, camera, cuts, rhythm, editing or continuation
<Audio N>
Audio used for voice timbre, music, rhythm, dialogue, sound texture or direct audio reuse
A reference file and a Subject are not the same thing. Picture 1 is the source image; Subject 1 can be the reusable person extracted from that image. One picture can define several subjects, and one subject can combine information from several references.
MiniMaxH4.org Reference Assignment Matrix
MiniMaxH4.org practical framework
Which reference role keeps what consistent
What Must Stay Consistent?
Best Reference Role
Face / identity
<Subject N> sourced from <Picture N>
Clothing
<Subject N>
Exact keyframe composition
<Picture N>
Body motion
<Video N>
Camera movement
<Video N>
Editing rhythm
<Video N>
Voice timbre
<Audio N>
Music style
<Audio N>
Whole original soundtrack
<Audio N> with explicit reuse instruction
Before writing an R2V prompt, assign every uploaded asset one primary job. A reference without a defined job adds context but not necessarily control.
Ask These 7 Questions Before Writing an R2V Prompt
Which reference defines the character's identity?
Which reference defines clothing or visual style?
Which reference defines the environment?
Which reference defines body motion?
Which reference defines camera movement?
Which reference defines voice?
Which reference defines music or audio continuity?
If two assets conflict, decide which one has priority before writing the final prompt.
Compact Reference-to-Video Prompt Skeleton
Starter Template
subject_definitions:
<Subject 1> is the main character whose facial identity and hairstyle come from <Picture 1>.
<Subject 2> is the black jacket and silver accessories shown in <Picture 2>.
<Video 1> provides the walking motion and slow handheld tracking-camera behavior.
<Audio 1> provides the voice timbre for <Subject 1> (S1).
summary:
[reference generation] Generate a new night-street scene that preserves <Subject 1>'s identity and <Subject 2>'s clothing while borrowing the walking motion and camera behavior from <Video 1> and the voice timbre from <Audio 1>.
retention_analysis:
<Subject 1>: preserve facial identity and hairstyle.
<Subject 2>: preserve clothing color, material and accessories.
<Video 1>: reference body motion and camera movement only.
<Audio 1>: reference voice timbre and delivery only.
detailed_description:
The target video uses a cinematic night-street style.
[Shot 1] <Subject 1> wearing <Subject 2> walks through a rain-soaked neon street. The walking rhythm and handheld tracking movement follow <Video 1>. <Subject 1> (S1) looks toward the camera and says: <d>[English] I knew you would find me.</d>
overall_soundscape:
Steady rain, wet footsteps and distant traffic.
non_diegetic_music:
Low electronic ambience with a slow restrained pulse.
This is a compact teaching example created by MiniMaxH4.org. Complex full-reference production prompts may require substantially more detail.
Reference Generation vs Video Editing vs Video Continuation
Reference task types
Task
Meaning
Reference generation
Borrow identity, style, motion, camera or audio traits without directly editing the source video
Video editing
Directly modify an existing source video
Video continuation
Generate new content that continues from a source video ending
Keyframe completion
Use an image as an exact first, intermediate or final frame anchor
Audio reuse
Reuse all or part of the original audio signal
Audio reference
Borrow characteristics such as voice timbre, rhythm or music style without copying the signal
Uploading a reference video does not automatically make the task “video editing.” If the source video only provides motion or camera behavior, the task is reference generation.
Better: Describe an observable shot, action and camera movement.
2. Trying to control resolution in the prose
Set resolution with the product or API controls. Use the prompt for audiovisual content and motion.
3. Too many camera movements in one shot
Choose one dominant movement unless the shot genuinely requires a compound move.
4. Re-describing the entire first image
In I2V, spend more prompt space on what changes after the first frame.
5. Treating first and last frames as two separate scenes
FL2V needs a believable transition path from the first state to the final state.
6. Uploading references without assigning roles
State whether each asset controls identity, style, motion, camera, voice or audio.
7. Mixing environmental sound and background score
Put physical and environmental sound in the scene or soundscape. Put audience-only score in non_diegetic_music.
MiniMax H4 Prompt Debugging Matrix
What to check first when a generation looks wrong
Problem in Output
First Prompt Element to Check
Wrong action
Action sequence in current Shot
Camera feels wrong
Camera movement terminology
Too many chaotic cuts
Shot count and timestamps
Character changes between shots
Subject description / reference assignment
I2V barely moves
Action after first-frame anchor
FL2V misses final frame
Transition path toward final state
Reference motion is ignored
Explicit <Video N> role
Voice identity is weak
<Audio N> assignment + speaker ID
Dialogue speaker changes
(S1)/(S2) consistency
Background music is wrong
non_diegetic_music
Environmental sound is missing
overall_soundscape
Change one layer at a time. If the camera is wrong, change the camera instruction before rewriting the character, lighting, dialogue and sound at the same time.
When Is a MiniMax H4 Prompt Too Complicated?
A longer prompt is useful only when the extra detail controls something visible or audible. Remove adjectives, repeated descriptions and duplicate instructions before removing actions, camera behavior, reference roles or audio timing.
MiniMaxH4.org simplification order
Remove first:
repeated adjectives
→ duplicate style descriptions
→ vague emotional language
→ redundant character description
Keep longest:
actions
→ state transitions
→ camera instructions
→ reference assignments
→ dialogue
→ synchronized sound
5 Copyable MiniMax H4 Templates
The five starter templates used on this page, gathered in one place. Each one is the same prompt shown in its section above — copy it, open the generator, and adapt it to your scene.
Text-to-Video Cinematic
Starter Template
A text-only scene with a clear visual start, a timed second shot, and full soundscape plus background music.
integrated_multimodal_description:
[Shot 1] Live-action cinematic night scene. A woman in a dark raincoat walks alone along a narrow train platform as heavy rain falls beyond the roof. A medium-wide tracking shot moves backward in front of her at walking speed. Reflections from blue platform lights move across the wet ground. Her shoes make small splashes as she walks.
[Shot 2] At 00:04.500, the camera cuts to a close-up from her left side. She stops, looks toward the approaching train and slowly closes her umbrella. White headlights grow brighter behind her as wind moves loose strands of her hair.
overall_soundscape:
Steady rainfall, wet footsteps, distant rail vibration, low station ambience and the rising metallic sound of an approaching train.
non_diegetic_music:
A restrained electronic pulse with sparse low synth notes, gradually increasing in intensity.
Start from the first frame, preserve what must stay fixed, and spend the prompt on what changes after it.
For the target video, at 0.00 seconds, <Picture 1> is the opening frame.
integrated_multimodal_description:
[Shot 1] Begin from <Picture 1> and preserve the character's face, clothing, color palette and initial composition. The woman slowly turns toward the window as wind begins moving the curtain behind her. The camera pushes in gently while sunlight shifts across her face. She lifts one hand and touches the glass.
overall_soundscape:
Quiet indoor room tone, soft fabric movement and distant city traffic through the window.
non_diegetic_music:
N/A
One continuous shot that explains the path from the opening frame to the target final frame.
How the reference pictures align with the target video:
Picture 1 is the opening frame and Picture 2 is the final frame.
integrated_multimodal_description:
[Shot 1] Begin from Picture 1. Preserve the woman's identity and clothing as she slowly walks from the center of the empty daylight street toward the camera. The camera pulls backward at the same speed. As she moves, shop lights switch on one by one, the sky darkens, reflections appear on the pavement and the scene gradually transforms into the neon night environment shown in Picture 2. Her movement remains continuous without a cut and the final composition settles precisely into Picture 2.
overall_soundscape:
Light footsteps, distant city traffic gradually becoming busier, soft electrical hum from illuminated signs.
non_diegetic_music:
Minimal electronic ambience slowly builds as the environment changes from day to night.
Two speakers with stable speaker IDs and spoken words kept inside the <d> tag.
The young woman with a quiet, breathy voice (S1) says:
<d>[English] We have five minutes before the storm reaches us.</d>
The older man with a low, rough voice (S2) replies:
<d>[English] Then we leave now.</d>
A compact six-section reference prompt that assigns each asset one explicit job.
subject_definitions:
<Subject 1> is the main character whose facial identity and hairstyle come from <Picture 1>.
<Subject 2> is the black jacket and silver accessories shown in <Picture 2>.
<Video 1> provides the walking motion and slow handheld tracking-camera behavior.
<Audio 1> provides the voice timbre for <Subject 1> (S1).
summary:
[reference generation] Generate a new night-street scene that preserves <Subject 1>'s identity and <Subject 2>'s clothing while borrowing the walking motion and camera behavior from <Video 1> and the voice timbre from <Audio 1>.
retention_analysis:
<Subject 1>: preserve facial identity and hairstyle.
<Subject 2>: preserve clothing color, material and accessories.
<Video 1>: reference body motion and camera movement only.
<Audio 1>: reference voice timbre and delivery only.
detailed_description:
The target video uses a cinematic night-street style.
[Shot 1] <Subject 1> wearing <Subject 2> walks through a rain-soaked neon street. The walking rhythm and handheld tracking movement follow <Video 1>. <Subject 1> (S1) looks toward the camera and says: <d>[English] I knew you would find me.</d>
overall_soundscape:
Steady rain, wet footsteps and distant traffic.
non_diegetic_music:
Low electronic ambience with a slow restrained pulse.
For T2V and keyframe-based H3 workflows, use a shot-by-shot integrated_multimodal_description plus overall_soundscape and non_diegetic_music. Describe visible actions, camera movement, dialogue and synchronized sound along the video timeline instead of relying on a flat list of style keywords.
How do I write camera movement in MiniMax H4?
Write the movement as part of the shot. Use precise terms such as push in, pull out, pan, truck, tilt, pedestal, arc or tracking shot, and add speed or amplitude only when those details matter.
How do I write dialogue in MiniMax H4?
Give each speaking character a stable ID such as (S1), then put the spoken words inside a <d>[Language] ...</d> block. Reuse the same speaker ID across later shots.
How should I prompt MiniMax H4 image-to-video?
Treat the image as the starting state. Preserve the visual details that must remain consistent, then spend the prompt on the action, camera movement, state changes and sound that occur after the first frame.
How should I prompt MiniMax H4 first-and-last-frame video?
Describe how the opening frame becomes the final frame. Focus on intermediate movement, pose changes, object changes, camera path, lighting and scene transitions rather than separately describing the two images.
How do MiniMax H4 references work?
Use Subject labels for reusable visual content, Picture labels for concrete image or keyframe anchors, Video labels for motion, camera, editing or continuation relationships, and Audio labels for voice, music, rhythm or audio reuse.
What is the difference between Picture and Subject in an H3 reference prompt?
Picture refers to the source image or a concrete frame anchor. Subject refers to reusable content extracted from references, such as a person, outfit, environment or style. A single Picture can provide multiple Subjects.
Does uploading a reference video mean I am editing that video?
No. If the video only supplies motion, camera movement, pacing or another reference characteristic, the task is reference generation. It becomes video editing when the source video itself is directly modified.
Where should background music go in a MiniMax H4 prompt?
Audience-only background score belongs in non_diegetic_music. Music that physically exists inside the scene, such as a radio, live instrument or club sound system, belongs in the shot description because the characters can hear it.
Official MiniMax H4 Prompt Sources
MiniMax H4 is developed by MiniMax. The primary prompt documentation below is published in the official MiniMax H4 model repository on Hugging Face.
Primary MiniMax documentation for T2VA, I2VA, FL2VA and L2VA prompt structure, shot timing, camera movement, dialogue, soundscape and background music.
Primary source for H3 task modes and model capabilities.
About the author
MiniMaxH4.org Editorial publishes hands-on AI video generation workflows, MiniMax H4 prompt engineering guides, and practical testing of multimodal video tools.
MiniMax H4 is developed by MiniMax. MiniMaxH4.org is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax.
