minimax-h3-prompt-workflow

MiniMax H3 Prompt and Workflow

Compile approved shot specs into MiniMax H3 / ComfyUI execution jobs. This skill is model-specific and must not alter narration meaning.

H3 constraints

• Native H3 canvas uses a 768px short edge and is capped around 768×1344, axes aligned to multiples of 32.
• At 24 fps, requested length snaps upward to the model's 17k+5 frame grid. Plan durations from legal frame counts rather than cutting arbitrary frames after generation whenever possible.
• Prefer the trained practical range for one generation unit; split long sequences into shots and chain only where continuity is meaningful.
• H3 is CFG-free in the local core workflow: use BasicGuider, not a negative-prompt CFG branch.
• First-frame I2V and first+last-frame FL2V are geometry/continuity controls, not substitutes for a clear motion prompt.

Mode selection

• T2V: new visual system or motivated hard reset; no exact composition to inherit.
• I2V: exact start composition/identity matters, destination can emerge naturally.
• FL2V: both starting state and destination state are semantically required and the in-between motion is physically plausible.
• R2V/Ref2VA: recurring identity/style/scene needs reference conditioning and the ref2va checkpoint is installed and validated.
• V2V/RV2V: source motion/timing itself is worth preserving; do not use merely to avoid designing a shot.

Prompt assembly

Write one independently executable prompt per generation unit:

1. minimal stable style/continuity baseline;
2. exact start state when not supplied by an image;
3. ordered visible events, each tied to a narration beat;
4. camera behavior with a reason;
5. required end state;
6. concise prohibitions only for known recurring failure modes.

Do not stuff the prompt with abstract goals such as “explain cognition clearly”. Translate them into observable geometry, action and state change.

Job manifest

For every render record:

• shot ID;
• narration time range;
• mode;
• width × height and provenance class (NATIVE, TEST, UPSCALED);
• legal frame count and expected duration;
• prompt;
• first/last/reference assets;
• seed, steps, sampler, scheduler;
• expected output prefix;
• semantic acceptance criterion;
• retake scope.