comfyui-workflows/custom/chris/minimax-h3-extender-working/prompt-set-specification.md

MiniMax H3 Prompt Set Specification

1. Purpose

This specification defines the standard structure for creating MiniMax H3 video prompt sets.

The format is designed around a practical production workflow:

1. Plan the complete sequence.
2. Define and create any required character, clothing, environment, prop, vehicle, or object references.
3. Allocate the available reference-image slots.
4. Create the Global Prompt.
5. Generate the initial video.
6. Feed the initial video into a chained JSON extension loop.
7. Carry forward continuity state explicitly between generated blocks.

The final deliverable should always be tight, clearly sectioned, actionable, and copy-ready.

---

2. Core Generation Model

2.1 Initial Video + Extension Chain

A prompt set consists of:

• An Initial Video Prompt
• A Global Prompt
• A sequence of JSON-defined video extensions

The initial generated video is fed into the first JSON extension. Each generated extension is then fed into the next.

Although the finished clips are stitched to appear as one continuous sequence, each generation block has no semantic memory of earlier prompts.

Therefore:

• The incoming video provides visual/start-frame continuity.
• The current prompt must provide semantic and state continuity.
• Persistent changes must be explicitly restated in later sections.

2.2 Duration

• 10 seconds is the maximum duration of a single generated block.
• Sections do not have to be 10 seconds long.
• Duration should match the natural length of the story beat.
• Short reactions, reveals, transitions, or simple actions may use shorter blocks.
• The workflow supports a custom per-section duration field.
• The exact JSON property for per-section duration is TBD and should not be invented until supplied.

2.3 Steps

Default:

"steps": 8

---

3. Reference Image System

MiniMax H3 supports a working bank of up to 7 reference-image slots for this workflow.

Reference allocation must be planned before video generation.

3.1 Character References

A consistent recurring character normally uses two images:

1. Primary full-body reference
2. Frontal facial reference

Therefore:

Consistent Characters Character Slots Used Slots Remaining

---

3 6 1 2 4 3 1 2 5

The actual number of remaining slots depends on whether an environment reference is used.

3.2 Maximum Consistent Characters

• Maximum: 3 consistent recurring characters per prompt set
• Each normally receives a full-body and facial reference.
• Additional incidental/background people may be generated when appropriate, but they should not be treated as persistent reference-driven characters unless explicitly planned.

---

4. Environment Modes

A prompt set can use either of two environment modes.

4.1 Referenced Environment

Use when:

• Spatial consistency matters.
• Characters interact with specific parts of the location.
• The sequence is relatively long.
• Fixed scene features are important to later actions.
• A distinctive location must remain visually consistent.

One image slot is assigned as the master environment reference.

Example:

<Picture 1> is the environment and master scene reference. Use it to establish the environment, spatial layout, lighting and overall scene.

The environment image must be designed around the complete planned story, not only the opening section.

If section 8 requires a gate, table, doorway, counter, caravan, staircase, etc., that feature should already be represented appropriately in the environment reference where practical.

4.2 Generated Environment

Use when precise location consistency is not important, particularly for:

• Short commercials
• Spoof adverts
• Short comedy sketches
• Simple studio scenes
• Quick standalone sequences
• Scenes where the environment is secondary to characters/products/actions

No environment reference image is required.

Instead:

• Describe the environment textually.
• Establish the broad scene in the Global Prompt.
• Add section-specific environmental details where required.
• Use the freed image slot for another useful reference asset.

---

5. Additional Reference Assets

Unused image slots can be assigned to:

• Products
• Props
• Vehicles
• Important objects
• Equipment
• Creatures
• Signs/graphics
• Other persistent visual elements

5.1 Composite Reference Images

A single image may contain multiple assets.

Example:

<Picture 7> is a prop reference sheet containing:
- the red holdall
- the silver cassette recorder
- the brass key

Treat these as three separate persistent story objects and preserve the appearance of each whenever shown.

Composite references must be defined before asset creation.

Every item should be:

• Clearly visible
• Visually separated where practical
• Explicitly named in the Global Prompt/reference mapping

---

6. Character Definition

Character planning should remain minimal during pre-production.

The purpose is to select an appropriate reusable character from the character bank and prepare their clothing.

Define only what is needed, such as:

• Name
• Approximate age, when relevant
• Gender/presentation, when relevant
• Very brief story role
• Physical characteristics only when critical to the story

Do not unnecessarily define:

• Race/ethnicity
• Hairstyle
• Body type

These should normally remain open so an appropriate existing character can be selected.

Character personality, relationships, mannerisms, accent, behaviour, and performance direction belong primarily in the Global Prompt.

---

7. Character Clothing Creation

Each recurring character should receive a highly descriptive re-clothing image prompt where new wardrobe is required.

The clothing prompt should preserve the existing character identity while changing the wardrobe.

Describe:

• Upper-body garments
• Lower-body garments
• Outerwear
• Footwear
• Accessories
• Materials
• Textures
• Colours
• Fit
• Layering
• Condition/wear
• Weather suitability
• Story-specific details

Avoid unnecessary physical identity descriptions.

Clothing should suit:

• Character role
• Scene
• Time period
• Weather
• Production style
• Story requirements

---

8. Global Prompt

The Global Prompt provides the shared semantic context for the complete prompt set.

It is currently required twice by the workflow because the workflow cannot automatically reuse the same value.

The two copies must be identical.

8.1 Global Prompt Contents

Include, where applicable:

• Broad environment/scene description
• Reference-image mapping
• Character identity mapping
• Additional asset mapping
• Composite-reference definitions
• Character personality
• Temperament
• Relationships
• Mannerisms
• Behaviour
• Voice/accent
• Speaking style
• Clothing identity
• General continuity rules
• Broad environmental behaviour
• Universal acting constraints
• Universal camera constraints
• Production style

8.2 Production Style

The Global Prompt should state the style type concisely, for example:

• Classic cinematic
• Modern cinematic
• 1990s sitcom
• Found footage
• Mockumentary
• Documentary
• Reality television
• Commercial/advertising
• Infomercial
• Home video/camcorder
• CCTV/surveillance
• Crime drama
• Horror
• Thriller

Do not overload the Global Prompt with detailed shot-by-shot cinematography.

Example:

Style: 1990s British sitcom with natural ensemble comedy and period-appropriate television presentation.

Specific camera instructions belong in individual video sections.

---

9. Camera Direction

Camera direction is defined primarily per section.

Each section may specify:

• Camera position
• Camera height
• Viewpoint
• POV
• Framing
• Subject focus
• Camera movement
• Reveal
• Transition
• Immediate view change
• Ending camera state

9.1 Useful Camera Position Bank

Available approaches include:

• Eye-level
• Chest-level
• Waist-level
• Ground-level
• Low-angle
• High-angle
• Overhead/top-down
• Elevated wide
• Distant observer
• Side profile
• Three-quarter front
• Three-quarter rear
• Direct frontal
• Direct rear
• Two-shot
• Three-shot
• Over-the-shoulder
• Reverse over-the-shoulder
• Dirty single
• Clean single
• Profile two-shot
• Across-table
• End-of-table
• Close follow
• Shoulder follow
• Side follow
• Front-facing follow
• Rear tracking
• Tabletop
• Floor camera
• Dashboard
• Back seat
• Doorway
• Window
• Inside cupboard/fridge/container
• Security camera
• Hidden observer

9.2 Framing

Useful framing includes:

• Extreme close-up
• Close-up
• Head-and-shoulders
• Medium close-up
• Medium shot
• Medium-wide
• Cowboy shot
• Three-quarter body
• Full body
• Wide
• Extreme wide

9.3 Movement

Useful movement includes:

• Locked/static
• Pan
• Tilt
• Push in
• Pull out
• Dolly
• Truck
• Tracking
• Follow
• Orbit
• Arc
• Rise
• Descend
• Fly-in
• Fly-out
• Whip-pan
• Handheld
• Gentle handheld
• POV head movement
• Follow eyeline
• Follow pointing gesture
• Turn toward sound
• Pull-back reveal

---

10. First-Person / POV

A section may switch into a character's first-person viewpoint.

Example:

Immediately begin from Bill's first-person point of view. The camera is positioned at Bill's eye level looking toward Nancy.

When using character POV:

• The POV character should normally not appear fully in front of the camera.
• Hands/arms/body may enter frame naturally.
• Reflections are an explicit exception.
• Other characters may look directly into the lens when they are actually looking at or speaking to the POV character.

This overrides a general "characters never look at camera" rule where logically necessary.

---

11. Chained Camera Continuity

A new JSON section is not automatically a new shot or camera cut.

By default, assume the next block continues from the incoming video's starting frame/state.

Use continuity-aware language such as:

Continuing naturally from the starting frame...

If the camera continues naturally:

Continuing from the starting frame, the camera remains beside Bill and begins moving slowly around him toward Nancy.

If the section must immediately begin from a different view, explicitly state this at the start:

Immediately change to Nancy's first-person point of view...

or:

Begin this section from a new low-angle position beside the table...

Do not assume MiniMax will infer a deliberate camera change simply because a new generation block has started.

---

12. Persistent Continuity State

Every generation block has to be told about important state changes that differ from the original references or original scene.

Example

If Bill's character reference shows him wearing a jacket and he removes it in section 3:

Section 4 and every relevant later section should state:

Bill had previously removed his jacket and remains without it.

Do not rely solely on the incoming video's final frame.

12.1 Track Persistent Changes To

• Clothing
• Accessories
• Hair/appearance changes caused during the story
• Wetness/dirt/damage
• Injuries or visible marks
• Held objects
• Prop ownership
• Object position
• Furniture position
• Open/closed doors
• Broken objects
• Vehicle state
• Drinks/food state
• Character position when spatially important
• Environmental changes
• Lighting/weather changes
• Any state that conflicts with the original reference image

12.2 Continuity Ledger

While designing the set, maintain a continuity ledger containing:

• Character clothing state
• Character position
• Character physical state
• Held objects
• Prop ownership
• Important object states
• Environment state
• Relevant emotional/story state
• Camera position
• Camera viewpoint

Each new section should be written from the current ledger.

---

13. Reference Priority

The intended priority is:

Reference Images

Define:

• Character identity
• Base character design
• Asset design
• Vehicle design
• Prop design
• Environment design where supplied

Current Video + Current Prompt

Define:

• Current state
• Current clothing changes
• Current object state
• Current position
• Current action
• Current environment changes
• Current camera state

If the story deliberately changes something from the original reference, the current prompt should explicitly reinforce that changed state.

---

14. Dialogue

For multi-character sequences:

• Explicitly name the speaker for each line.
• Keep dialogue appropriate to the available block duration.
• State important pauses/reactions where required.
• Specify silence when incorrect dialogue assignment is a realistic risk.
• Preserve established accents and speaking styles through the Global Prompt.

Example:

Bill leans forward and asks, "Do you think it's time?"

Becca exhales before replying, "I don't know. We normally wait till sunset."

---

15. Initial Video Prompt

The Initial Video Prompt establishes the first generated block.

It should contain:

• Starting composition
• Character positions
• Initial actions
• Dialogue
• Body language
• Camera position
• Camera movement
• Environmental activity
• Relevant sound/ambience
• Ending action/state that leads naturally into extension 01

The Global Prompt is supplied alongside the Initial Video Prompt.

---

16. JSON Extension Structure

Baseline structure:

{
 "defaults": {
 "duration_seconds": 10,
 "steps": 8
 },
 "shots": [
 {
 "id": "existing_video_extension_01",
 "prompt": [
 "Continuing naturally from the starting frame..."
 ],
 "seed": "SEED_VALUE"
 }
 ],
 "prompt_prefix": [
 "GLOBAL/REFERENCE INSTRUCTIONS"
 ]
}

16.1 Extension IDs

Use sequential IDs:

existing_video_extension_01
existing_video_extension_02
existing_video_extension_03
...

16.2 Number of Sections

The number of extensions depends on the requested story.

Ten extensions may be used for a longer set, but this is not a fixed requirement unless requested.

16.3 Per-Section Duration

The workflow supports custom duration per extension.

Exact JSON syntax: TBD.

Until supplied, do not invent the property.

---

17. Prompt Prefix

The JSON prompt_prefix contains shared reference/global instructions required by the extension loop.

Example with three characters:

"prompt_prefix": [
 "<Picture 1> is the environment and master scene reference, use this to establish the environment, spatial layout, lighting and overall scene.",
 "",
 "<Picture 2> is the primary full-body reference for Bill, use this as the primary reference for Bill's body proportions, overall appearance, hair and clothing.",
 "<Picture 3> is a frontal facial reference for Bill, use this as a complementary reference for Bill's facial identity.",
 "",
 "<Picture 4> is the primary full-body reference for Nancy, use this as the primary reference for Nancy's body proportions, overall appearance, hair and clothing.",
 "<Picture 5> is a frontal facial reference for Nancy, use this as a complementary reference for Nancy's facial identity.",
 "",
 "<Picture 6> is the primary full-body reference for Becca, use this as the primary reference for Becca's body proportions, overall appearance, hair and clothing.",
 "<Picture 7> is a frontal facial reference for Becca, use this as a complementary reference for Becca's facial identity."
]

The actual mapping must reflect the reference plan for the current production.

---

18. Required Deliverable Order

Every complete prompt-set response should be kept concise and delivered in this order.

PREP

1. Scene Overview

• Premise
• Broad story direction
• Environment mode
• Character count
• Relevant time/weather/setting information

2. Character Selection

• Minimal information required to select reusable characters

3. Re-Clothing Prompts

• One copy-ready image prompt per character

4. Environment Creation

• Only when using a referenced environment
• One copy-ready environment image prompt

5. Extra Asset Creation

• Props
• Vehicles
• Products
• Composite reference sheets
• Other required visual references

6. Reference Allocation

• Exact Picture 1--7 mapping
• Clearly state unused slots if any

VIDEO GENERATION

7. Global Prompt

• Copy-ready
• Contains global character/story/reference/style rules

8. Initial Video Prompt

• Copy-ready
• First generated block

9. JSON Extension Set

• Copy-ready
• Sequential chained continuation prompts
• Persistent state restated as required

10. Global Prompt --- Workflow Copy

• Exact duplicate of section 7

---

19. Planning Rule

The complete story should be planned before finalising the asset requirements.

This ensures:

• The environment contains required features.
• Important props are created before generation.
• Vehicles receive reference slots when necessary.
• Composite sheets are planned efficiently.
• Reference slots are not wasted.
• Later sections do not suddenly require unavailable visual references.

The user-facing deliverable should remain concise even though the complete sequence has been planned internally.

---

20. Output Philosophy

A MiniMax H3 prompt set should be:

• Production-ready
• Concise
• Clearly sectioned
• Easy to follow top-to-bottom
• Explicit about creation jobs
• Explicit about reference allocation
• Copy-ready
• Continuity-aware
• Camera-aware
• Efficient with reference slots
• Designed around MiniMax H3's lack of semantic memory between generated clips

Avoid:

• Unnecessary explanation
• Repeating the same instructions outside required workflow duplication
• Overloading the Global Prompt with shot-specific camera directions
• Assuming previous prompt context is remembered
• Assuming a new extension automatically creates a camera cut
• Introducing important assets after reference planning is complete

---

21. Current Defaults

---

Setting Default

---

Maximum block duration 10 seconds

Steps 8

Maximum consistent characters 3

Character reference images 2 per consistent character

Total reference slots 7

Environment reference Optional

Camera direction Per section

Production style Concise definition in Global Prompt

Persistent changed state Restated in every relevant subsequent block

Per-section duration JSON field TBD

---

22. Outstanding Workflow Detail

The following remains to be added once supplied:

• Exact JSON property and syntax for custom per-section duration.

Once known, it should become part of the canonical JSON extension specification.