Add MiniMax H3 Ref2VA prompt guide

2026-08-10 19:28:56 +00:00
parent c985203c58
commit ae16868927
2 changed files with 397 additions and 0 deletions
@@ -0,0 +1,389 @@
 # MiniMax H3 Ref2VA Prompt Guidelines
 Source basis: Chris-provided `MiniMax H3 Ref2VA Reference & Scene Guidelines — V2`
 ## Purpose
 Practical prompt and reference-structure guide for `MiniMax H3 Ref2VA`,
 especially on an `8GB VRAM` workflow where the safe planning baseline is:
 - `480p`
 - maximum `10s` per generated shot
 - one coherent shot per generation
 - explicit roles for every reference image
 ## Core rule
 Reduce the system to five responsibilities:
 1. `Master scene`
   Defines where the shot happens and how it initially looks.
 2. `Full-body references`
   Define who is present and their overall appearance.
 3. `Headshot references`
   Reinforce whose face it is.
 4. `Specialist references`
   Define clothing, props, vehicles, or other narrow details.
 5. `Prompt`
   Defines what happens during the next `10` seconds.
 ## Reference hierarchy
 Use this authority order:
 1. `master scene / environment`
 2. `primary full-body character references`
 3. `facial identity references`
 4. `clothing references`
 5. `props / vehicles / specialist references`
 6. `text prompt`
 Within a character:
 - `full body` controls overall appearance, body proportions, hair, and clothing
 - `headshots` refine facial identity
 ## Master scene rules
 Prefer `<Picture 1>` as the master scene when possible.
 It should control:
 - environment
 - composition
 - camera position
 - perspective
 - field of view
 - lighting
 - initial subject placement
 - important object placement
 Preferred phrasing:
 `<Picture 1> is the master scene, environment and composition reference.`
 `Use <Picture 1> as the sole visual reference for the environment.`
 `Preserve its initial camera position, framing, perspective, field of view, horizon, lighting, environment geometry, object positions, proportions and spatial layout.`
 `Do not duplicate, offset, mirror, layer or reconstruct a second version of the environment.`
 ## Environment rules
 If a reference already establishes the environment, do not re-describe every
 visible object in text unless something changes.
 Prefer:
 `The environment, background, spatial layout and lighting remain as established by <Picture 1>.`
 Over a full textual rebuild of the room.
 ## Character reference grouping
 When multiple images represent the same person, explicitly say so.
 Preferred phrasing:
 `<Picture 2>, <Picture 3> and <Picture 4> all represent Subject 1.`
 `Use their consistent identity information together to preserve Subject 1 throughout the video.`
 `Do not describe them as separate subjects.`
 ## Full-body reference role
 Primary use:
 - overall appearance
 - body proportions
 - build
 - hair
 - clothing
 - silhouette
 Preferred phrasing:
 `<Picture 2> is the primary full-body reference for Subject 1.`
 `Use <Picture 2> to define Subject 1's overall appearance, body proportions, build, hair and clothing.`
 `Ignore the background, environment, camera composition and location shown in <Picture 2>.`
 ## Headshot reference role
 Primary use:
 - facial identity
 - facial structure
 - eyes
 - nose
 - mouth
 - skin appearance
 Preferred phrasing:
 `<Picture 3> is a facial identity reference for Subject 1.`
 `Use <Picture 3> to reinforce Subject 1's face, facial structure, eyes, nose, mouth, skin appearance and other facial characteristics.`
 `Do not use <Picture 3> to determine body proportions, clothing, environment or scene composition.`
 ## Full-body vs headshot conflict rule
 Always resolve authority explicitly:
 `<Picture 2> defines Subject 1's overall body, proportions, hair and clothing.`
 `<Picture 3> and <Picture 4> provide additional facial identity information for the same Subject 1.`
 `Use the headshot references to improve facial identity while preserving the body, clothing and overall appearance established by <Picture 2>.`
 ## Multi-character separation
 Keep groups explicit to avoid identity leakage:
 `<Picture 2> and <Picture 3> represent Subject 1.`
 `<Picture 4> and <Picture 5> represent Subject 2.`
 `Do not allow the identity or physical characteristics of one subject to influence another subject.`
 ## Background isolation rule
 Character references often contain irrelevant backgrounds. State that those
 backgrounds should be ignored.
 Preferred phrasing:
 `Use all character reference images only for their assigned character information.`
 `Ignore their original backgrounds, locations, scene compositions, camera positions and environmental lighting.`
 `The environment is defined exclusively by <Picture 1>.`
 ## Specialist references
 Narrow responsibility prevents conflicts.
 Example clothing phrasing:
 `<Picture 5> defines the clothing worn by Subject 1.`
 `Use <Picture 5> only as a clothing reference.`
 `Ignore the person, face, body, pose, environment and background shown in <Picture 5>.`
 Example vehicle phrasing:
 `<Picture 6> defines the appearance of the vehicle.`
 `Use <Picture 6> only as a visual reference for the vehicle.`
 `Ignore its original environment, camera position and background.`
 ## 8GB VRAM constraints
 Assume this baseline unless direct testing proves otherwise:
 - resolution: `480p`
 - maximum duration: `10s`
 - one coherent action beat per shot
 Do not design a single generation as a long sequence of unrelated events.
 ## Action design for 10 seconds
 Good:
 - `Subject 1 walks towards Subject 2.`
 - `Subject 2 turns towards Subject 1.`
 - `They briefly look at each other.`
 - `The camera slowly tracks forward.`
 Bad:
 - entering a room
 - crossing the room
 - talking
 - sitting
 - picking something up
 - finishing a conversation
 - standing
 - leaving
 all in one `10s` shot
 ## Timing guidance
 Simple timing is useful when sequencing matters:
 - `0-3 seconds`: approach
 - `3-7 seconds`: reaction
 - `7-10 seconds`: settle / hold
 Avoid over-micro-timing unless necessary.
 ## Dialogue guidance
 Dialogue must fit within the same `10s` budget as motion and camera changes.
 If the line is long, keep action simple.
 Preferred pattern:
 - character holds position or makes one simple move
 - one spoken line
 - one visible reaction
 - one restrained camera movement
 ## Camera guidance
 The master scene defines the opening frame.
 Prompt text should describe how that camera changes, not replace it with a new
 simultaneous framing.
 Good camera phrasing:
 - `The camera slowly pushes forward.`
 - `The camera gently pans right to follow Subject 1.`
 - `The camera slowly arcs around both subjects while maintaining their spatial relationship.`
 ## Expression guidance
 Headshots lock identity, not frozen expression.
 Expression belongs in performance instructions:
 `Subject 1 initially appears relaxed.`
 `As Subject 2 speaks, their expression gradually becomes concerned.`
 ## Recommended prompt architecture
 Use this section order:
 1. `Reference definitions`
 2. `Scene anchor`
 3. `Reference restrictions`
 4. `Subject placement`
 5. `Action / performance`
 6. `Camera`
 7. `Dialogue / audio`
 ## Single-character template
 ```text
 <Picture 1> is the master scene, environment and composition reference.
 <Picture 2>, <Picture 3> and <Picture 4> all represent Subject 1.
 <Picture 2> is Subject 1's primary full-body reference and defines
 their overall appearance, body proportions, build, hair and clothing.
 <Picture 3> and <Picture 4> are complementary facial identity
 references for Subject 1.
 Use these headshot references to reinforce Subject 1's facial identity
 while preserving the overall appearance established by <Picture 2>.
 Use <Picture 1> as the sole visual reference for the environment.
 Preserve its initial composition, camera position, perspective,
 field of view, lighting, environment geometry and spatial layout.
 Ignore the backgrounds, environments and compositions shown in
 <Picture 2>, <Picture 3> and <Picture 4>.
 [ACTION — MAXIMUM 10 SECONDS]
 [CAMERA]
 [DIALOGUE]
 [AUDIO]
 ```
 ## Multi-character template
 ```text
 <Picture 1> is the master scene, environment and composition reference.
 <Picture 2>, <Picture 3> and <Picture 4> all represent Subject 1.
 <Picture 2> is the primary full-body reference for Subject 1
 and defines their overall appearance, body proportions, build,
 hair and clothing.
 <Picture 3> and <Picture 4> are complementary facial identity
 references for Subject 1 and provide additional information about
 their facial structure and appearance.
 <Picture 5> and <Picture 6> both represent Subject 2.
 <Picture 5> is the primary full-body reference for Subject 2
 and defines their overall appearance, body proportions, build,
 hair and clothing.
 <Picture 6> is a facial identity reference for Subject 2.
 The opening frame uses the environment and composition established
 by <Picture 1>.
 Use <Picture 1> as the sole visual reference for the environment.
 Preserve its initial camera position, framing, perspective,
 field of view, horizon, lighting, environment geometry,
 object positions, proportions and spatial layout.
 Do not duplicate, offset, mirror, layer or reconstruct a second
 version of the environment.
 Use the full-body character references to establish the overall
 appearance of their respective subjects.
 Use the headshot references to reinforce facial identity while
 preserving the body and overall appearance established by the
 corresponding full-body reference.
 Ignore the backgrounds, environments, camera positions and scene
 compositions shown in all character reference images.
 Do not allow the identity or physical characteristics of one
 subject to influence another subject.
 [ACTION AND PERFORMANCE — MAXIMUM 10-SECOND SHOT]
 [CAMERA MOVEMENT]
 [DIALOGUE]
 [AUDIO / ENVIRONMENTAL SOUND]
 ```
 ## Efficiency rule
 Do not add references just because they exist.
 Good default set for one important character:
 - `1 x` full-body reference
 - `1 x` frontal headshot
 - `1 x` three-quarter headshot
 The goal is the smallest reference set that clearly defines the required
 information.
 ## Practical default checklist
 - work at `480p`
 - treat `10s` as the safe maximum shot duration
 - use one coherent action or interaction per generation
 - use a master scene whenever possible
 - give the master scene authority over environment and composition
 - give each important character one primary full-body reference
 - add headshots only when facial identity needs reinforcement
 - explicitly state which images belong to the same character
 - stop character-reference backgrounds from influencing the master scene
 - use the text prompt mainly for action, performance, camera, dialogue, and audio
@@ -28,6 +28,14 @@
 - Keep `optimized/text-to-video/minimax-h3-3070-turbo-master/` as the first MiniMax default.
 - Keep the INT8 reference-video workflow as the next donor for future low-VRAM optimization.
 ## Ref2VA prompt guidance
 - For `MiniMax H3 Ref2VA`, treat `480p` and `10s` as the safe planning baseline on `8GB VRAM` unless direct tests prove otherwise.
 - Use one master-scene reference to control environment/composition, one primary full-body reference per subject to control overall appearance, and headshots only to reinforce facial identity.
 - Explicitly state when multiple images represent the same subject, and explicitly tell the model to ignore character-reference backgrounds so they do not fight the master environment.
 - Use the prompt mainly for action, performance, camera movement, dialogue, and audio rather than re-describing scene details already visible in the references.
 - Full guidance now lives in `optimized/notes/minimax-h3-ref2va-prompt-guidelines.md`.
 ## What needs Monday proof
 - whether the Turbo LoRA workflow is just technically possible or actually usable