MiniMax-H3 Crossover Scene Rules

Production-tested rules for writing multi-character crossover scenes (characters from different shows and films sharing one story) as independent MiniMax-H3 T2VA clips. These sit on top of the official base guide: the official guide owns field names and grammar; this document owns the four-section crossover layout and the anti-artifact rules learned from rendering hundreds of these scenes.

1. The four-section scene layout

Every scene is one plain-text prompt in exactly this order, sections separated by one blank line, no markdown, no bold, no bullets:

subject_definitions:
<Subject 1> [Character] (played by [Actor]) from [Show]
<Subject 2> [Character] (played by [Actor]) from [Show]

integrated_multimodal_description:
[Shot 1] [Visual style, lighting, time of day, setting]. A single medium close-up of <Subject 1> [Character] (S1) alone in frame, wearing [wardrobe anchors]. <Subject 2> is not in frame. The shot opens with <Subject 1> [Character] (S1) already speaking, with no silent establishing beat. The camera holds a static shot as <Subject 1> [Character] (S1) says: <d>[English in [Character]'s voice from [Show]] Dialogue.</d>

[Shot 2] At 00:05.500, the shot cuts to a single medium close-up of <Subject 2> [Character] (S2) alone in frame, wearing [wardrobe anchors]. <Subject 1> is not in frame. The camera pushes in with small amplitude at slow speed as <Subject 2> [Character] (S2) [physical action] and says: <d>[English in [Character]'s voice from [Show]] Dialogue.</d>

overall_soundscape:
[Room tone, physical sounds, ambient environment, non-verbal human sounds. Never voices.]

non_diegetic_music:
N/A

subject_definitions: lists every character who appears on screen in this scene, one per line, in the exact form <Subject N> Character (played by Actor) from Show. Use the character, actor and show strings you were given verbatim - never rename, never substitute another actor. If the show string carries extra look notes after a comma, keep them as wardrobe/physical anchors in the shots but put only the show name after from.
Never bold section headers. subject_definitions: not **subject_definitions:**.
[Shot 1] never carries a timecode. [Shot 2] and later always open with At MM:SS.mmm, the shot cuts to ....
Scenes with no people (never the case here unless asked) write subject_definitions: then N/A on its own line.

2. Speaker binding - the rules that stop face and voice blending

The primary speaker of the scene is always <Subject 1>. H3 binds dialogue audio to <Subject 1>. Reorder the cast per scene so the character with the most spoken dialogue is <Subject 1>.
Every spoken line is wrapped: <d>[English in Character's voice from Show] words</d>. Speaker phrase, ID (S1), action and delivery stay outside the tag; only the language tag and the words go inside.
Stable IDs (S1), (S2), (S3) across every shot of the scene. (S1,S2) only for a single word or short phrase spoken in unison.
Silence mandates for everyone on screen who is not speaking: <Subject 2> Name (S2) remains completely silent with his mouth closed and lips sealed, speaking no dialogue. Add it whenever a non-speaker is in frame, and after a speaker finishes early in a shot: then falls silent with her mouth closed and lips sealed, speaking no further dialogue for the remainder of the shot.
Isolated single-subject shots are the safe default for dialogue exchanges: a single medium close-up of <Subject 1> Name (S1) alone in frame ... <Subject 2> is not in frame. Say who is not in frame explicitly.
Shared frames are allowed and encouraged for variety (establishing shots, walk-and-talks, handoffs, toasts, reactions to the same event) with safeguards: state each subject's position (<Subject 1> stands screen-left, <Subject 2> stands roughly three feet to his right at screen-right, separated by a clear gap with no overlap), one active speaker at a time, silence mandate on the others, restate wardrobe anchors, and never more than 3 people visible in one shot (prefer 2).
After an action beat, cut to the victor alone before they deliver a line, with an off-camera eye-line (looking down at the fallen guard off-camera).
Off-screen voiceover: says in an off-screen voiceover: <d>...</d> while his lips remain completely closed.
Never name a voice or dialogue in overall_soundscape: (and Saul's fast-talking voice is a known double-voice bug). Non-verbal sounds (a cough, a laugh) are fine. Never put musical score in the soundscape; it belongs in non_diegetic_music:.

3. Identity anchoring

Restate key wardrobe / physical anchors for every subject in every shot they appear in (brown leather jacket over a dark flannel shirt, silver pendant necklace; bald head, facial scar over his right eye). Generic terms (the agent, in a shirt) drift.
Anchor era/age for long-career actors when it matters: John McClane (played by 35-year-old Bruce Willis, 1988 Die Hard era) from Die Hard, and restate the age-defining traits in the shots.
Keep costume, hair, props and demographics identical across every scene of the story.
Visible on-screen text goes in English double quotes: a red neon sign reading "QUARANTINE".

4. Pacing, duration and dead air

Each scene lands between 5 and 20 seconds; prefer 12-15 s for dialogue scenes, 5-8 s for transitions, punches and arrivals. Never exceed 20 s.
Fill the whole duration with dialogue and physical action. No multi-second silent stares, no frozen smiles at the end. If dialogue ends 1-2 s early, end on concrete movement: pocketing a prop, turning and walking away, stepping out of frame, closing a case, sliding into a car.
Proven structures for a 15 s scene:
- A: S1 speaks (00:00-00:07) -> S2 speaks and acts to the end (00:07-00:15)
- B: S1 speaks (00:00-00:06) -> S2 speaks (00:06-00:11) -> S2 pockets prop, turns, walks (00:11-00:15)
- C: S1 (00:00-00:05) -> S2 (00:05-00:10) -> S1 quick closer while stepping forward (00:10-00:15)
Spread cut timecodes across the actual scene duration (00:05.500, 00:10.500 for 15 s; 00:04.000, 00:08.000 for 12 s). Vary the shot pattern between scenes.
Open T2V clips on the speaker already talking: The shot opens with <Subject 1> Name (S1) already speaking, with no silent establishing beat.

5. Staging, story and flow across scenes

Show, don't tell. Characters physically do things: uncork a flask, unlatch a case, tap a glowing interface, unhinge a trunk, aim a device, lean on a hood, push through a door. Never a static monologue into the lens.
Cause and effect on screen: light from a portal reflects on wardrobe; a thrown knife hits, the target groans and drops the case - impact and reaction in the same shot.
Precise action mechanics: executes a razor-sharp tactical drop, ducking under the gunfire in one fluid motion and sliding into a low crouch behind the concrete pillar on the left of frame beats he dodges.
Grounded entrances: every character has an on-screen reason to be there (a badge, a job, a line, an invitation, a delivery). Stagger arrivals; do not dump the whole cast into scene 1.
Story blocks: keep the same 2-3 characters for 2-4 consecutive scenes resolving a beat before moving to another group. Adjacent scenes share connective tissue: end on a look off-screen, a handoff, a walk toward a spot, a question - and start the next scene from that beat. Keep key props alive across scenes.
Location variety: 2-3 scenes in one spot, then move (lobby -> corridor -> office -> parking lot). Every shot states style, lighting temperature and time of day, and these stay identical inside a sequence unless the story jumps time.
Exits and entrances: anchor the camera position and the movement vector (looking back at the glass entrance from outside ... they walk out through the doors, moving away from the building behind them).
Camera vocabulary: [Motion] + [with small/large amplitude] + [at slow/fast speed] - push in, pull out, pan, truck, tilt, pedestal, arc, tracking, static, POV, roll, shake.
Play every character in character: their vocabulary, rhythm, humour and moral compass from their own show, colliding with the others'. That collision is the comedy or drama.
Skip fiddly held instruments (a guitar in hands smears); a cup, a file, a phone, a badge, a weapon at rest are safe.

8.4 KiB Raw Permalink Blame History