Pika H3 Flex Prompt Guide
Official syntax schemas, reference marker formatting, advisory camera rules, and copyable prompt templates for T2VA, FL2VA, and Ref2VA pipelines.
All prompt templates provided below are original examples created independently to illustrate the structural field names documented in Pika's official prompt-guidance.md. They do not copy vendor marketing showcase prompts. Furthermore, these examples are documentation-derived and untested by us on physical hardware; we make no claims of verified rendering success or benchmark accuracy.
Four Essential Prompting Principles
Pika's documentation highlights four operational principles governing how the underlying transformer interprets cinematic direction:
1. Camera Moves are Advisory
Directives such as "slow pan right", "forward dolly", or "tracking shot" act as soft directional vectors. To prevent indefinite camera panning or drifting, specify a narrative resting event (e.g., "...pan right until the character steps into the doorway and turns toward the light").
2. Avoid Sub-Second Cut Demands
Demanding exact micro-cuts at specific fractions of a second (such as "at 02.4s, hard cut to...") is unreliable. The temporal diffusion schedule exhibits an adherence variance of approximately 0.2 to 0.3 seconds. Structure sequences in natural chronological narrative blocks (e.g., 00:00–00:03 followed by 00:03–00:06).
3. Treat Framing as Spatial Description
Isolated tags like "wide shot" or "extreme wide" frequently default to medium framing. Instead, describe the physical environment relative to the subject: "The lone climber appears small in the lower third of the frame beneath towering granite cliffs that occupy the upper two-thirds."
4. Synchronized Audio & Dialogue Syntax
Audio is generated natively alongside visual frames. Spoken lines should be wrapped in quotation marks or custom tags (<d>...</d>) and anchored to visible facial movements. Conclude prompts with dedicated overall_soundscape: and non_diegetic_music: blocks.
Official Structural Fields vs. Optional Prose
It is important to distinguish official parser fields recognized by the conditioning pipeline from optional prose descriptions:
| Field / Marker | Status | Workflow | Technical Purpose |
|---|---|---|---|
integrated_multimodal_description: |
Official Field | T2VA, FL2VA | Primary narrative block integrating visual style, pacing, and chronological actions. |
overall_soundscape: |
Official Field | All | Specifies environmental diegetic sound effects, foley, and spoken dialogue cues. |
non_diegetic_music: |
Official Field | All | Defines musical genre, tempo, instruments, and mood. |
subject_definitions: |
Official Field | Ref2VA | Explicitly binds attached references (<Picture 1>, <Video 1>, <Audio 1>) to identity labels. |
summary: |
Official Field | Ref2VA | High-level synopsis of the intended scene and reference task. |
retention_analysis: |
Official Field | Ref2VA | Mandatory persistence rules (e.g. keeping facial structure, wardrobe, or color palettes identical). |
detailed_description: |
Official Field | Ref2VA | Chronological shot-by-shot timeline for multi-reference generations. |
| "Cinematic 35mm film grain" | Optional Prose | All | Descriptive aesthetic keywords included within the narrative description. |
Original Prompt Templates by Task
1. Text-to-Video & Audio (T2VA)
Uses the official three-field schema for unconditioned text-to-video generation:
integrated_multimodal_description: 35mm anamorphic aesthetic with moody warm tungsten street lighting. 00:00–00:03 — Low-angle tracking shot glides backward along rain-slicked asphalt as an archivist in a wet trench coat strides under a buzzing diner sign. 00:03–00:06 — Camera decelerates and pans slightly right as the archivist pauses before a brass doorway, unlatches a vintage briefcase, and looks upward as rain streams down their face. overall_soundscape: Steady rhythmic rain pattering against pavement; low rolling thunder in the distance; metallic snap of briefcase latches opening; muffled footsteps in puddles. non_diegetic_music: Melancholic noir cello melody with low muted brass chords and gentle brush snare rhythm.
integrated_multimodal_description:, overall_soundscape:, and non_diegetic_music: fields. Untested by us.
2. First/Last-Frame to Video & Audio (FL2VA)
Uses the official keyframe alignment syntax mapping input image tags to timestamps:
Use <Picture 1> as the frame at 0.00s and <Picture 2> as the frame at 5.00s. integrated_multimodal_description: Soft natural morning light streaming through an art studio window with shallow depth of field. 00:00–00:03 — Beginning from the potter seated at the wheel in <Picture 1>, the camera gently dollies forward as their clay-coated hands shape a spinning wet earthen cylinder. 00:03–00:05 — The camera arcs smoothly upward to eye level, resolving naturally into the final composition of <Picture 2> as the artisan smiles at the completed vessel. overall_soundscape: Continuous motorized whir of the pottery wheel; wet clay sloshing softly under palm pressure; gentle studio room acoustics. non_diegetic_music: Warm acoustic fingerstyle guitar with light hall reverberation.
Use <Picture 1> as the frame at 0.00s...) followed by the standard multimodal description and audio fields. Untested by us.
3. Reference-to-Video & Audio (Ref2VA)
Uses the full six-section format documented for multimodal referencing with up to 12 inputs:
subject_definitions: <Picture 1> is the lead botanist wearing tortoiseshell glasses and a dark green utility vest; <Picture 2> is the interior greenhouse environment featuring misting nozzles and lush tropical palms. summary: The lead botanist inspects an exotic flowering orchid inside the humid greenhouse, speaking diagnostic observations aloud. retention_analysis: Retain the facial bone structure, hairstyle, glasses, and vest styling from <Picture 1> exactly across all frames. Maintain the architectural lighting and plant density from <Picture 2>. detailed_description: 00:00–00:03 — Medium tracking shot moves slowly left around <Picture 1> as they gently support a blooming orchid stem with gloved fingers. 00:03–00:06 — The botanist turns slightly toward the camera and speaks clearly: "Specimen respiration is nominal. Growth rates match baseline." Spoken lip kinematics match phonemes accurately. overall_soundscape: Soft hissing mist from automated greenhouse overhead nozzles; ambient moisture drips; clear spoken dialogue from <Picture 1>; gentle rustle of palm fronds. non_diegetic_music: Subtle ambient synthesizer pads with airy flute accents.
subject_definitions:, summary:, retention_analysis:, detailed_description:, overall_soundscape:, non_diegetic_music:) matching official prompt-guidance.md specifications. Untested by us.