01
Playback order beats keyword clouds
If the model must choose between five aesthetics and three camera moves, the shot usually collapses. Sequence the decisions the viewer will actually see.
MiniMax H3 prompt guideT2VA · I2VA · First/last frame · Omni reference
Write MiniMax H3 prompts like compact shot production briefs
A useful MiniMax H3 prompt is more than a mood board. Give one shot a clear subject, a readable setting, a controlled movement, a camera path, optional sound intent, and an ending—then remove anything that competes with the main action.
Precise prompt formula
Subject + scene + movement + camera + sound intent + ending
Write decisions in playback order. One dominant movement and one clear camera path usually communicate more than a stack of conflicting instructions.
- Subject
- Who or what matters
- Scene
- Where the shot lives
- Movement
- What changes over time
- Camera + end
- How it is seen and resolves
- T2VA
- Text-led scene invention
- I2VA
- First-frame motion
- FL2V
- Opening → ending
- R2VA
- Omni references
- Prompt limit
- 7,000 characters
- Duration
- 4–15 seconds
Short version
A strong MiniMax H3 prompt is a timeline, not a tag list
MiniMax H3 can invent a full scene, animate a locked frame, bridge first and last images, or follow omni references. The safest way to guide that mix is a short production brief in playback order: scene, subject, action, camera, sound intent, ending.
02
Mode first, adjectives second
Text-only asks H3 to invent the world. First-frame asks it to move a known world. Reference mode asks it to inherit signals. Write for the job you selected.
Templates
Start from the right MiniMax H3 template, then remove anything unnecessary
The content can look similar, but the job is different. Pick the mode before polishing language.
Scene-building template
Use when MiniMax H3 must invent subject, environment, and opening composition from text. Ratio is required in pure text mode.
T2VA template
[Duration and aspect intent]. [Subject and defining appearance] in [specific scene and time]. [Primary action or change over the clip]. [Camera path and framing evolution]. [Lighting, palette, material, atmosphere]. [Optional dialogue or sound intent]. End on [final composition or emotional beat].Motion-first template
Use when the first frame already establishes identity, composition, and the initial aesthetic.
I2VA template
The [main subject visible in the image] [performs one clear action]. [Environment or secondary elements change]. [Camera path] while preserving the subject’s defining appearance and the original visual style. Optional sound: [voice/ambience cue]. End with [final pose or framing].Transition template
Describe how the first controlled image evolves into the last rather than re-describing both frames in isolation.
FL2V template
Beginning from the first frame, [subject action and environmental transition]. [Camera movement and timing]. Preserve [identity, object, or style constraints] as the composition evolves naturally into the supplied final frame.Reference map template
Map each reference before restating style adjectives. Do not mix reference roles with first/last-frame roles.
R2VA template
Use reference image(s) for [character / product / style]. Use reference video for [motion / camera / rhythm]. Use reference audio for [voice timbre / delivery] if provided. Create a [duration] shot in [scene]. Subject action: [one readable action]. Camera: [path]. Keep referenced identity consistent. End on [final beat].Prompt formula
Build the MiniMax H3 brief from six decisions
Add detail only when it earns control. Extra adjectives are not free—they compete with motion.
01
Subject
Name the main person, product, creature, or form. Add appearance details only when they affect recognition.
02
Scene space
Place the subject in a concrete environment: night street, studio, forest edge, bridge interior, stylized world.
03
Movement or change
Describe the action and how the scene evolves. Video needs a change over time, not a list of static visual tags.
04
Camera path
Choose framing and travel: static, track, push, pull, rise, pan, orbit, or a restrained combination in sequence.
05
Sound intent
If dialogue, room tone, or voice reference matters, state it. Pair audio references with visual media in r2va mode.
06
Final beat
State where the shot should settle: a close expression, a wide reveal, a completed turn, or a still product frame.
Duration is part of the prompt design
MiniMax H3 accepts 4–15 integer seconds. A five-second product reveal and a twelve-second character beat need different action density. Do not pack a three-act story into one short clip.
Input modes
Choose the prompt mode before writing the first line
The words can look similar, but MiniMax H3’s job changes with the media roles you attach.
T2VA
Text invents the whole frame
Spend tokens on appearance, environment, and opening composition. Ratio cannot be left adaptive in pure text generation.
I2VA
First frame owns identity
Do not re-describe every pixel. Direct motion, camera, light change, and what must stay consistent.
FL2V
Bridge two compositions
Write the transition logic: what moves, what morphs, what the camera does between the supplied ends.
R2VA
References own the signals
Map which asset controls face, wardrobe, motion, camera language, or voice. Avoid first/last-frame roles in the same request.
Reference checklist
Describe references like a director, not a file dump
Uploading media is not enough. The prompt should assign each reference a job.
Identity
Which image owns the face, product, or logo?
Say it explicitly when multiple images could compete.
Motion
Which video owns body language or camera rhythm?
Keep reference clips short and focused—documented limits are 2–15s per clip.
Voice
Which audio owns timbre or delivery?
Audio cannot travel alone; pair it with image or video references.
Exclusion
Are you mixing first/last frames with reference roles?
If yes, split into two tasks. The public contract treats those modes as mutually exclusive.
Budget
Are you under 12 media files and 7,000 prompt characters?
Trim low-value adjectives before trimming the reference map.
Ending
Can you picture the final frame in one sentence?
If not, the motion often drifts instead of resolving.
Layer timeline
Put action, camera, and sound where they happen
Time words help MiniMax H3 understand priority. Use them to stage the shot, not to micromanage every frame.
Opening
Establish space and subject
First second: who we are watching and where we are. Avoid starting mid-chaos unless the first frame already holds the read.
Development
One primary action
Middle window: the movement that justifies generating video instead of a still.
Camera
One path, maybe a second beat
If you need two camera ideas, sequence them—track then push—rather than stacking them as equal priorities.
Exit
Resolve the composition
Final beat: expression settle, product still, wide reveal, or completed gesture.
Example prompts
Four reusable MiniMax H3 shot shapes
These are newly written structural examples. Swap subject and setting while preserving hierarchy.
Rain-lit cyclist
One subject action, one camera path, and a final emotional beat.
5–8s cinematic shot
Create an 8-second 16:9 cinematic night shot. A courier in a reflective jacket cycles through a rain-lit city intersection. Water sprays from the tires as neon reflections stretch across the road. The camera tracks beside the bicycle, then gently closes on the courier’s focused expression. Cool blue practicals with restrained magenta accents, wet-surface realism. Soft tire hiss and distant traffic. End on a readable close three-quarter face.Watch reveal
Let the first frame own the product design while the prompt directs light and camera motion.
Controlled image animation
The watch remains centered as a narrow band of warm light travels across the brushed metal case. Fine mist moves slowly behind the product. Camera trucks slightly right while preserving watch geometry and the dark marble surface. No redesign of the dial. End on a clean three-quarter product angle with the dial fully readable.Studio to street bridge
Describe the evolution between two controlled compositions.
Transition brief
Beginning from the first frame, the model walks forward as studio softboxes dissolve into rainy street practicals. Fabric motion stays natural and identity remains consistent. Camera slowly pulls back while the environment completes the transition into the supplied final frame. End exactly on that last composition with settled posture.Character with voice cue
Map references before restating style adjectives.
R2VA brief
Use reference images for the character’s face and wardrobe. Use reference video for relaxed walking rhythm. Use reference audio for warm mid voice timbre. Create a 10-second medium shot on a windy coastal boardwalk at golden hour. The character walks toward camera, smiles once, and says a short free-spirited line. Camera tracks backward smoothly. Keep identity and wardrobe consistent. End on a stable medium close-up.Common failure patterns
If the output feels messy, simplify the brief first
Most weak MiniMax H3 prompts do not need more words. They need a clearer subject, action, mode, or camera priority.
Keyword cloud
Static style tags with no change over time
Replace “cinematic, beautiful, 4K, dramatic” with a subject performing one visible action inside a defined scene.
Mode mismatch
Writing t2va language for an i2va job
If the first frame already exists, stop rebuilding wardrobe and background from scratch unless change is intentional.
Reference dump
Many files, no assignment
Map which asset controls face, motion, or voice. Unmapped references become noise.
Competing cameras
Several moves at equal priority
Choose one dominant path. If a second is necessary, place it later as a sequence.
Overloaded story
Too many beats for one clip
Give the shot one job. Build longer sequences from several generated clips.
Illegal mix
First/last frames plus reference roles
Split the work. The public MiniMax-H3 contract treats those mode families as mutually exclusive.
Before generation
A six-question MiniMax H3 prompt check
Read the prompt once as a director and once as a timeline.
Mode
Is the generation mode explicit?
Text-only, first frame, first+last, or omni reference—pick one contract.
Subject
Is one subject clearly dominant?
A viewer should know what to watch before the camera starts moving.
Scene
Is the environment concrete?
Name place, time, weather, or production context when it changes the visual result.
Movement
Does something visibly change?
Describe an action, transformation, expression, or environmental event across time.
Camera
Is there one primary camera path?
Sequence secondary moves; do not stack contradictions.
Ending
Can you picture the final frame?
A clear last beat helps motion resolve instead of drifting.
Camera vocabulary appendix
Useful camera language when you need explicit control
Natural-language camera direction is enough for many MiniMax H3 shots. When precision matters, keep commands sparse and sequential.
Truck
truck lefttruck rightMove the camera laterally while preserving viewing direction.
Pan
pan leftpan rightRotate left or right from a relatively fixed position.
Push and pull
push inpull outPhysically move toward or away from the subject.
Pedestal / rise
riselowerRaise or lower the camera without forcing a tilt.
Tilt
tilt uptilt downRotate the view upward or downward.
Track / follow
tracking shotfollowMove with the subject so environment changes while framing stays readable.
Zoom
zoom inzoom outChange apparent focal length rather than traveling through space.
Static
static shotlocked offHold the camera steady and let performance carry the clip.
Handheld feel
subtle handheldrestrained shakeUse sparingly so subject motion remains readable.
Prompt questions
Direct the important part first
Should I write a long MiniMax H3 prompt?
Length is not the goal. Use enough detail to identify subject, setting, movement, camera path, optional sound intent, and ending—then remove instructions that compete with the primary action. The documented limit is 7,000 characters.
How is image-to-video prompting different from text-to-video?
The first frame already establishes subject, composition, and style. Focus the prompt on motion, camera behavior, environmental change, and the features that must remain consistent.
How should I write omni-reference prompts?
Map each reference to a job—face, wardrobe, motion, camera rhythm, or voice—before adding style adjectives. Keep total media within documented caps and never mix reference roles with first/last-frame roles.
Can I request dialogue or voice?
You can state spoken intent in the prompt, and reference audio can guide timbre when paired with visual media in reference mode. Keep speech short enough for the chosen duration.
Where do I learn the exact API fields?
Use the MiniMax H3 API guide on this site and MiniMax’s official V2 documentation for content roles, duration, ratio, and status handling.
Write one clear MiniMax H3 shot. Then test the direction.
Start with the mode, add one useful camera decision, and compare the result before increasing complexity.