I previously rushed through a test and published an article right away just to ride the traffic around Minimax H3. I was wrong—I shouldn't have written such a perfunctory piece of filler content. Minimax H3 has strict requirements for prompt formatting, and prompts cannot be written using conventional thinking. So I carefully studied the official prompting guide and wrote this tutorial.
>
Let's begin
>
When writing video prompts, many people are used to cramming the characters, setting, action, camera, and style into a single paragraph: a girl sits in a café, a dog lunges toward the cookie in her hand, the scene is warm and funny with a cinematic feel, the camera pushes in…
This approach is not completely unusable, but when you upload a character image, a setting image, an action video, and an audio reference at the same time, the model can easily become confused:
Which image is responsible for the character's appearance; which image defines the setting; which video is only for action reference; which audio must be preserved exactly; which elements must remain consistent; and which elements may be changed?
Seedance2.0's solution for full-reference mode is to use the @asset syntax.
MiniMax H3's solution for full-reference mode is instead to write the prompt as a clearly structured video production brief.
The official requirements state that a complete prompt should consist of six sections, with the main body written in English. Only character dialogue, lyrics, and text that actually appears in the frame may remain in the original language.
>1. First understand one core principle
>
A prompt for full-reference mode is not answering: What story does this video tell?
It is answering: What content should the model extract from which assets, and at what time, in which shot, and where in the frame should that content be used?
So don't write only a plot summary: a girl meets a dog in a café.
This was the mistake I made when I rushed my test. I was also producing filler content just to chase traffic.
Instead, write visuals that the model can execute:
Medium shot. The girl sits on an orange sofa on the right side of the frame, holding a cookie in her right hand. A young man enters from the left, leading a white Samoyed. The dog suddenly moves toward the cookie, pulling the leash taut, and the girl quickly draws her hand back.
The former is a story; the latter is a video prompt.
A qualified shot description should explain at least: shot size and composition + character appearance and position + environment and lighting + a single action + state change + camera movement + sound and dialogue.
This is actually not much different from the prompt structure used by most video models; it is essentially a general-purpose structure.
However, Ha螺's official guide gives a particularly important reminder:
>detailed_description must not be reduced to a plot summary or a simple reference-asset mapping table. It must clearly state what can actually be seen and heard in every shot.
>
>2. The six sections of a complete prompt
>
The standard structure for MiniMax H3 full-reference mode is as follows:
1. subject_definitions
Defines the content in the reference assets that needs to be identified and tracked separately.
2. summary
Briefly explains what video should be generated and what role each reference asset plays.
3. retention_analysis
Tells the model which features in the reference assets must be preserved completely, which only need to be partially referenced, and which features should be transferred to other subjects.
4. detailed_description
The most important section. In playback order, describe the visuals, actions, camera, sound, and dialogue shot by shot.
5. overall_soundscape
Describes ambient sound and sounds that can be heard in the real world, such as wind, footsteps, and indoor background noise.
6. non_diegetic_music
Describes background music that only the audience can hear and that the characters in the scene cannot hear.
These six sections must be arranged in a fixed order.
>3. How to use the four reference tags
>
The most important skill in full-reference mode is labeling reference content correctly.
The official guide defines four tags:
<SubjectN>
<PictureN>
<VideoN>
<AudioN>
N is numbered starting from 1.
<SubjectN>: content that should actually appear in the video
A Subject is not a “file,” but a visual element extracted from a file that needs to be used in the finished video. It can be a character; animal; object; setting; clothing; prop; special effect; visual style; action; expression; or pose.
For example, if you upload an image of a young wolf demon character, your prompt should look like this:
<Subject1>is the young wolf demon in<Picture1>,with grey fur,drooping ears,worn brown clothes,and anervous expression.
This means: Subject1 is the young wolf demon in Picture1, with grey fur, drooping ears, worn brown clothing, and a nervous expression.
One image can provide multiple Subjects.
For example, a café photo can provide both:
<Subject1>is the young woman in<Picture1>.
<Subject2>is the coffee-shop environment in<Picture1>.
The same Subject can also come from multiple assets at the same time:
<Subject1>is the woman whose appearance comes from<Picture1>and whose walking motion comes from<Video1>.
This means: <Subject1> is a woman whose appearance comes from <Picture1> and whose walking motion comes from <Video1>.
In other words: Picture1 is responsible for the character's appearance; Video1 is responsible for the character's walking motion.
This is one of the most practical capabilities of Minimax H3 full-reference mode: combining attributes from different assets onto the same target character.
<PictureN>: a specific visual anchor
In Minimax H3, uploading an image does not necessarily mean that you must define a separate <PictureN>.
Only when the image itself serves one of the following purposes should it be separately defined as a Picture:
the video's first frame; the video's last frame; a keyframe; a shot-composition reference; a storyboard image; or a frame that needs to be edited directly.
For example:
<Picture2>is the first frame of[Shot1],showing the wolf demon standing outside there cruitment hall.
This means: <Picture2> is the first frame of [Shot1], showing the wolf demon standing outside the recruitment hall.
If an image is only used to establish a character's appearance, you do not need to write a separate Picture line; simply include it in the Subject definition:
<Subject1>is the wolf demon in<Picture1>.
The most common mistake is to upload five images and mechanically define five Pictures.
The correct approach is to first determine for yourself: is this image a specific visual frame, or is it only providing a character, clothing, setting, or style?
Use Subject when it only provides features.
Use Picture only when it directly controls a specific frame or shot composition.
<VideoN>: the structural source of an entire video
Video is mainly used to describe the relationship between an entire reference video and the target video, for example:
editing the original video; continuing the original video; referencing the original video's camera movement; referencing its editing rhythm; referencing its shot order; or referencing its overall temporal structure.
For example:
<Video1>is the source video for the target video edit.
This means: <Video1> is the source video for editing the target video.
However, if you only want to use the movement of a person in the reference video, that person or movement should still be defined as a Subject:
<Subject2>is the nervous hand-rubbing motion from<Video1>.
This means: <Subject2> is the nervous hand-rubbing motion from <Video1>.
Do not define everything in a video as <Video1> simply because the source asset is a video.
You can think of it this way: <VideoN> governs the structure and temporal relationship of the entire video; <SubjectN> governs the specific content extracted from the video for use.
<AudioN>: audio assets
Audio can be used to:
copy an entire audio track exactly; copy part of it; reference a character's vocal timbre; reference speaking rhythm; reference emotion and tone; reference music style; reference sound-effect qualities; or continue the original audio's beat.
For example:
<Audio1>is the voice-timbre reference for<Subject1>(S1).
This means Audio1 is only responsible for Subject1's speaking voice timbre.
Note: the fact that an uploaded video file contains sound does not mean that you must automatically define an Audio.
Define <AudioN> only when the sound genuinely needs to be copied or referenced.
Also, Video and Audio are numbered independently; their numbers do not need to correspond one-to-one.
>4. First determine which type of task you have
>
Does this tutorial make you feel that everything was easier before reading it, and that afterward you've become completely lost?
It is actually simple.
At the beginning of summary, you need to use square brackets to indicate the task type.
There are six common types:
1. keyframe completion
Reference images are used directly as the first frame, last frame, or a keyframe.
[keyframe completion]
2. reference generation
Reference character, setting, style, action, camera movement, or audio characteristics.
But do not directly edit or continue the original video.
[reference generation]
This is also the most commonly used type for ordinary multimodal reference generation.
3. video editing
Directly modify an existing video.
[video editing]
4. video continuation
Continue generating from the end of the original video.
[video continuation]
5. audio reuse
Directly copy the original audio signal.
[audio reuse]
6. audio reference
Do not copy the original sound; only reference its timbre, rhythm, tone, or music style.
[audio reference]
A single task can involve multiple types at the same time:
[video continuation+keyframe completion+audio reference]
Do not write video editing simply because you uploaded a video.
If you are only referencing the video's camera movement or rhythm, it usually still falls under reference generation.
Only directly modifying the original video qualifies as video editing.
Only continuing from the end of the original video qualifies as video continuation.
>5. How to tell the model “what must not be changed”
>
retention_analysis is used to clarify how much of the reference material should be preserved.
Visual materials use the following four relationships:
1, fully_preserved
Preserve completely. Applicable when: the same character must be retained; the clothing cannot change; the scene layout cannot change; the first or last frame must match strictly.
<Subject1>:fully_preserved the character's grey fur,drooping ears,brown clothes,and facial identity areretained.
This means that the character's grey fur, drooping ears, brown clothes, and facial features in <Subject1> are all preserved.
2, partially_preserved
Preserve part of it, while allowing modifications.
For example: preserve the character's identity but change the clothing; preserve the scene structure but change the time and lighting; preserve the main action but change the speed.
3, attribute_transfer
Transfer reference attributes to another object.
For example: transfer an adult's movements from the reference video to a young wolf demon; put the clothing from the reference image on another character; transfer a particular expression to the target character.
4, weak_reference
Preserve only a general resemblance. Applicable when: referencing only the atmosphere; referencing only the visual style; referencing only the general composition; referencing only the shot pacing.
If you upload audio rather than video, use:
fully_copy: copy the entire content;
partially_copy: copy part of the content;
reference: reference only the timbre, rhythm, or style;
weak_reference: preserve only the general atmosphere.
This section is not asking you to evaluate generation quality. It tells the model in advance which reference relationships are strong constraints and which are merely weak references.
>Six. What really determines the result is detailed_description
>
detailed_description is the main body of the prompt.
The official recommendation for video generation is to write a prompt of approximately 350 to 500 English words, but do not pad the word count mechanically. For videos with a lot of dialogue, first ensure that the dialogue fits naturally within the available time.
For video editing tasks, adjust it according to the complexity of the original video.
1. Establish the overall style in one or two sentences first.
Put the style description before the first shot:
The target video uses a stylized 3D animated mockumentary style, with slightly exaggerated expressions, warm daylight, and dry absurdist comedy.
This means: the target video uses a stylized 3D animated mockumentary style, with slightly exaggerated expressions, warm daylight, and dry absurdist comedy.
Do not pile up dozens of conflicting style terms, such as:
cinematic, anime, photorealistic, watercolor, clay animation, cyberpunk, Chinese ink painting
Cinematic, anime, photorealistic, watercolor, clay animation, cyberpunk, and Chinese ink painting.
The correct phrasing should be:
A stylized 3D animated mockumentary with rough hand-painted textures and natural daylight.
Choose one primary style, then add one or two texture characteristics.
2. Do not write a time for the first shot.
[Shot1]
Mark the transition time for subsequent shots:
[Shot2]At 00:03.000,...
[Shot3]At 00:06.500,...
The timestamp indicates when the new shot begins.
3. Write each shot in a fixed order.
Beginners can use the following order to effectively avoid common pitfalls:
Step 1: Shot size and composition
A medium shot shows...
A close-up frames...
A wide shot establishes...
Step 2: Character position
<Subject 1> stands on the left side of the frame.
Step 3: Character appearance
When the character appears for the first time, briefly restate the most important identifying features:
<Subject 1>, the grey-furred young wolf demon in worn brown clothes, ...
There is no need to fully redefine the character in subsequent shots.
Step 4: A single action
He slowly hands the recommendation letter to the interviewer.
Do not arrange multiple actions in one short shot, such as entering, sitting down, handing over materials, turning his head, falling, standing up, speaking, and pushing the door open to leave.
The model recognizes these actions, but it has difficulty arranging their sequence correctly within a few seconds.
A more reliable approach is to split them into separate shots. This is also a commonly used method for creating multiple actions in other models:
Shot 1: Walk into the room;
Shot 2: Sit down;
Shot 3: Hand over the recommendation letter;
Shot 4: The interviewer looks down to examine it.
Step 5: Changes in expression and state
Do not write only “nervous”; describe visible changes:
His ears droop slightly, his shoulders tighten, and he avoids eye contact.
Step 6: Camera movement
The camera slowly pushes in.
The camera makes a slight handheld pan to the right.
The camera remains static.
When necessary, specify the speed and magnitude. When no camera movement is needed, simply describe a static shot.
Step 7: Sound and dialogue
Specify the sound source, manner of speaking, and exact dialogue.
>Seven. The standard format for dialogue
>
Every person who actually speaks needs a fixed speaker ID:
(S1)
(S2)
(S3)
The IDs are assigned according to the order in which the characters speak for the first time and must not be changed arbitrarily afterward.
For example:
<Subject 1> (S1) says in a nervous young male voice, <d>[Chinese] 我有三封推荐信，两封是我自己写的。</d>
Here:
<Subject1> represents the reference character in the frame;
(S1) represents the source of the voice;
[Chinese] indicates the language of the dialogue;
<d>...</d> encloses the complete dialogue.
When the same character speaks again later, continue using (S1).
When the character speaks off-screen:
<Subject 1> (S1), speaking off-screen, says...
Note:
Dialogue, narration, and lyrics should all be written inside <d>; the original language should be retained and does not need to be translated into English; write [unclear] for unintelligible original dialogue and do not guess; if you are referencing only the timbre, do not bring the original dialogue from the reference audio into the new video; dialogue should end with a normal period, question mark, or exclamation mark.
The official system treats “visual subjects” and “voice sources” as two separate tracking systems: Subject identifies who the character is, while S1/S2 identify who the voice comes from. It is quite cumbersome.
>Eight. Environmental sound and background music must be separated
>
1, overall_soundscape
Describe sounds that actually exist in the world shown on screen.
overall_soundscape: Quiet indoor room tone, faint paper rustling, wooden chair creaks, and distant crowd noise from outside the recruitment hall.
Overall soundscape: quiet indoor room tone, faint paper rustling, wooden chair creaks, and distant crowd noise from outside the recruitment hall.
Sounds that actually exist in the on-screen world include wind, rain, street noise, indoor background noise, footsteps, doors opening, clothing rustling, and objects colliding.
2, non_diegetic_music
Describe music that only the audience can hear.
non_diegetic_music: A restrained plucked-string comedy score at a medium-slow tempo, with brief pauses before each punchline.
When using BGM, it is best to clearly specify: which instruments are used; the musical tempo; the mood; changes in volume; whether it gradually builds; and whether it stops before the punchline.
Do not repeat complete dialogue in these two sections. Dialogue belongs only in detailed_description.
>Nine. The eight most common mistakes beginners make
>
Mistake 1: Defining every reference file as Picture
When an image is responsible only for the character's appearance, define it as Subject rather than defining it separately as Picture.
Mistake 2: Writing the person in a reference video as Video
Characters, objects, and actions belong to Subject. Video mainly represents the editing, continuation, pacing, or shot structure of the entire video.
Mistake 3: Writing only “reference a certain image”
The model does not know what exactly you want to reference.
Incorrect: Use Picture 1 as reference.
Correct: <Subject 1>'s facial identity, grey fur, drooping ears, and worn brown clothes come from <Picture 1>.
Mistake 4: Assigning too many vague tasks to one reference asset
Do not write: Use Video1 for everything.
Instead, separate all reference elements: the character's movements come from Video1; Video1 is a weak reference for editing rhythm; the audio is not copied; the scene comes from Picture2.
Mistake 5: Writing only the plot without describing the visuals
A prompt such as “He attends an absurd interview” cannot be executed directly by the model.
You must specify: where he is standing; where the interviewer is; who moves first; the magnitude of the movement; whether the camera moves; and when the dialogue occurs.
Mistake 6: Packing too many actions into one shot
One shot should handle one action. For complex events, split them into multiple shots.
Mistake 7: Repeatedly changing character IDs
Once a character is (S1), it must remain (S1) throughout the entire video.
Mistake 8: Confusing “copying the voice” with “referencing the voice”
To retain the original audio track, write audio reuse.
To reference only the timbre or speaking style, write audior eference.
>Ten. A complete template beginners can use directly
>
This is probably the section beginners like most 😂
>
subject_definitions:
<Subject1>is[the person, animal, object, or scene to be generated],whose[appearance, clothing, identity, and other features]come from<Picture1>.<Subject2>is[the second visual subject],whose[action, expression, or appearance]come from<Video1>.<Picture2>is the[first frame/keyframe/last frame/composition reference]of[ShotN].<Video1>is the source video used as a reference for[camera movement/editing rhythm/temporal structure].<Audio1>is the voice-timbre reference for<Subject1>(S1).summary:[reference generation+audio reference]The target video shows[characters, setting, and core event].<Picture2>defines[the specific visual function].<Video1>provides[the video reference function].<Audio1>provides[the audio reference function].retention_analysis:<Subject1>(appears in[Shot1],[Shot2]):fully_preserved[features that must be retained].<Subject2>(appears in[Shot2],[Shot3]):partially_preserved[what is retained and what may change].<Picture2>([Shot1]first frame):fully_preserved[composition and positions that must be strictly retained].<Video1>(camera movement and pacing):weak_reference[content that only needs to be referenced approximately].<Audio1>:reference[only reference the timbre, emotion, and speaking rhythm; do not copy the original signal].detailed_description:The target video uses a[primary style]with[lighting, colors, and overall texture].[Shot1]A[shot size]shows[environment and composition].<Subject1>,[key appearance features],stands or sitsat[screen position].[The character performs a primary action].The camera[ camera movement]. [Environmental sounds].[Shot2]At 00:03.000,the shotcuts to a[shot size]of<Subject2>.[Changes in the subject's position, action, and state].<Subject1>(S1)says in a[timbre, emotion, speed],<d>[Chinese]specific dialogue.</d>[Shot3]At 00:06.000,the shotcuts to[a new composition].[The character performs a primary action].<Subject2>(S2)replies in a[voice description],<d>[Chinese]specific dialogue.</d>The camera[ camera movement or remains static].The shot ends on[the final visual state].overall_soundscape:[ambient background noise, footsteps, clothing rustle, object impacts, and other realistic sounds].non_diegetic_music:[instruments, rhythm, mood, volume, and how the music develops].
>
>Eleven. Complete example: The Young Wolf Demon Attends an Interview at Huaguo Mountain
>
Suppose the following have been uploaded:
Picture1: character image of the young wolf demon;
Picture2: Huaguo Mountain recruitment hall;
Video1: a video of a job applicant nervously handing over a résumé;
Audio1: voice reference for the young wolf demon.
It can be written as follows:
subject_definitions:
<Subject 1> is the young wolf demon in <Picture 1>, with grey fur, drooping ears, a thin body, worn brown clothes, and a permanently nervous expression.
<Subject 2> is the monkey interviewer, wearing a red official robe and sitting behind a large wooden recruitment desk.
<Subject 3> is the recruitment-hall environment in <Picture 2>, featuring stone walls, red banners, crowded wooden benches, and a large sign reading “Huaguo Mountain Talent Recruitment Office”.
<Subject 4> is the nervous document-handing motion from <Video 1>, transferred to <Subject 1>.
<Audio 1> is the voice-timbre reference for <Subject 1> (S1), providing a young male voice with a cautious and slightly trembling delivery.
summary:
[reference generation + audio reference] The target video shows <Subject 1> attending a job interview with <Subject 2> inside <Subject 3>. The nervous document-handing action of <Subject 4> is transferred to <Subject 1>. <Audio 1> provides the voice-timbre reference for <Subject 1>. The scene uses dry absurdist comedy and ends with the interviewer rejecting the résumé because it lacks a sufficiently powerful recommendation letter.
retention_analysis:
<Subject 1> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the wolf demon's facial identity, grey fur, drooping ears, thin body, worn brown clothes, and nervous appearance are retained.
<Subject 2> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the monkey interviewer's red official robe, seated posture, and bureaucratic expression are retained.
<Subject 3> (appears in [Shot 1], [Shot 2], [Shot 3]): fully_preserved - the stone recruitment hall, wooden desk, red banners, benches, and visible Chinese recruitment sign are retained.
<Subject 4> (appears in [Shot 1]): attribute_transfer - the cautious document-handing motion is transferred to <Subject 1>.
<Audio 1>: reference - the target speaker follows the youthful, cautious, slightly trembling vocal quality without copying the original audio signal.
detailed_description:
The target video uses a stylized 3D animated mockumentary look with rough hand-painted textures, warm daylight entering through the stone doorway, slightly exaggerated facial expressions, and dry absurdist comedy.
[Shot 1] A medium-wide static shot establishes <Subject 3>, the crowded stone recruitment hall with red banners, wooden benches, and the large sign reading “Huaguo Mountain Talent Recruitment Office”. <Subject 2>, the monkey interviewer in a red official robe, sits behind a large wooden desk on the right side of the frame. <Subject 1>, the thin grey-furred wolf demon in worn brown clothes, stands on the left. Using the nervous document-handing motion transferred from <Subject 4>, he slowly extends a folded résumé toward the interviewer with both hands. His ears droop and his shoulders remain tense. Papers rustle softly as the résumé reaches the desk.
[Shot 2] At 00:03.000, the shot cuts to a close-up of <Subject 2>. He lowers his eyes toward the résumé without touching it. His expression remains serious and slightly impatient. <Subject 2> (S2) asks in a flat bureaucratic voice, <d>[Chinese] Do you have the Great Sage's letter of recommendation?</d> He raises one eyebrow and taps the desk once with his index finger. The camera remains static.
[Shot 3] At 00:06.000, the shot cuts to a close-up of <Subject 1>. He tightens his grip on the résumé, lowers his head, and answers using the cautious, slightly trembling voice referenced from <Audio 1>. <Subject 1> (S1) says, <d>[Chinese] No, but I can handle hard work.</d> After a short pause, the shot cuts back slightly wider as <Subject 2> pushes the résumé back across the desk with one finger and says, <d>[Chinese] We don't lack hard work here; we lack connections.</d> <Subject 1> remains frozen with both hands extended as the final frame holds for one second.
overall_soundscape:
A low indoor crowd murmur continues throughout the recruitment hall. Soft paper rustling, wooden chair creaks, distant footsteps, and one dry finger tap on the desk remain clearly audible.
non_diegetic_music:
A restrained plucked-string comedy score at a medium-slow tempo plays quietly beneath the scene. The music stops briefly before the interviewer's final line, followed by one short wooden-percussion note during the final frozen frame.
In other words, young one, if you want to use MiniMax H3 effectively, you have to work in English.
But it's not a big problem. You can write the Chinese version in the required format first and then translate it. However, the fastest way is to give the complete example above to an AI and ask it to imitate the style, then replace the elements afterward 😂
>12. Finally, remember this five-step method
>
When facing any complex reference material, don't start by writing a long prompt.
Complete these five steps first:
Step 1: Break down the material
What does each piece of material provide: characters, setting, action, composition, camera movement, pacing, or sound?
Step 2: Define the tags
Use Subject for reusable visual content, Picture for specific visual anchors, Video for relationships across an entire video, and Audio for sound.
Step 3: Decide how much to preserve
Which elements must be preserved completely, which may be modified, and which are only weak references?
Step 4: Break down the shots
Following the actual playback order, break complex events into multiple shots, each containing a single action.
Step 5: Add sound
Distinguish between character dialogue, ambient sound, and background music that only the audience can hear.
The underlying logic of MiniMax H3's full-reference prompts can be condensed into one sentence:
Don't tell the model what files you uploaded; tell it what role each piece of material serves and how it should take effect in which shot of the final video.
MiniMax H3 has strict formatting requirements, completely different from models such as Seedance 2.0. For now, following this format is essential for producing a good video.
If you don't want that much hassle, wait a little longer. Someone will optimize this process.
>That's it.
end