Pika H3 Flex Prompt Guide: Camera, References and Sound

A useful video prompt identifies the subject, one main action, camera behavior, visual treatment and sound. For Pika H3 Flex, reference roles also matter: say which input establishes identity, which establishes the product or setting, and which motion should change.

Reviewed November 9, 2026. The templates below are original suggested starting points, not results we have benchmarked on Flex. This independent website currently provides MiniMax H3 and other listed models, not Flex itself. Do not treat a template as an API request schema.

Start with a shot brief

Write the following before adding adjectives:

• Subject and retained details: who or what is visible, and which details must stay recognizable?
• Action: what changes during this clip? Give the action a natural stopping point.
• Camera: fixed, push-in, tracking, or a restrained orbit? Pick one dominant move first.
• Look: framing, lighting, materials and a consistent visual style.
• Sound: dialogue, location sound, effects and music. State when something should be absent.

ComfyUI’s official guide recommends explicit shot and sound descriptions. Its full-reference form uses fields including subject_definitions, summary, retention_analysis, detailed_description, overall_soundscape and non_diegetic_music, with reference labels such as <Picture 1>. Follow the exact format of your selected official workflow or endpoint.

Six adaptable templates

1. Product image to a restrained ad shot

Input: your own clear product image. Keep packaging fully visible. Suggested starting point: a short image-to-video clip at a supported lower resolution; choose the publishing ratio before generating.

Use the supplied product image as the visual anchor. Keep the bottle shape,
cap, label placement and material recognizable. The bottle stays upright on
a stone plinth. A soft reflection moves across the glass while the camera
makes one slow push-in and stops before the label fills the frame.
Soft side light, warm neutral background, no new text or extra objects.
Sound: gentle room ambience, no speech and no music.

Inspect the label, silhouette and reflections. If lettering changes, reduce movement and camera rotation; do not assume a stronger identity instruction guarantees exact text.

2. Portrait micro-expression

Input: a permitted portrait with clear eyes and an unobstructed face.

The person blinks once, breathes naturally and gives a small relaxed smile.
Keep the head position and hairstyle stable. The camera remains fixed.
Soft window light and a quiet background. No scene change, speech or music.

Check eyes, teeth and face shape. If identity drifts, remove the camera move before adding further expression instructions. The image-to-video walkthroughs include an existing portrait inspiration clip and explain the limits of its historical metadata.

3. Camera move with an endpoint

A ceramic cup sits on a large empty wooden table. Keep the cup small enough
that the table and surrounding room remain visible. The camera slowly arcs
a short distance to the right, then comes to rest with the cup still centered.
Morning side light, realistic ceramic glaze, quiet room tone, no music.

Describe visible composition, not only “wide shot.” Treat a requested angle or timestamp as direction rather than guaranteed precision.

4. One speaker, one line

Medium shot of one adult presenter in a quiet studio. The presenter looks
at the camera and says exactly: "Here is the first idea."
Natural speaking pace and restrained gestures. The camera stays fixed.
Sound: one clear voice and soft room tone. No other speakers, music or subtitles.

Use a model that supports audio. Keep dialogue short enough for the clip, and check pronunciation and synchronization. A silent-output model cannot acquire audio support from a prompt.

5. Consistent stylized motion

A paper-cut fox crosses a layered forest from left to right, pauses beneath
a tree, and looks up. Keep the same flat paper texture, warm limited palette
and simple silhouette throughout. One continuous side-on shot, no style change.
Sound: soft paper rustle and forest ambience, no speech or score.

Decide whether style or physical realism matters more. Avoid combining incompatible media instructions unless the transition itself is the goal.

6. Separate reference roles

Inputs: two licensed images of the same character from compatible angles, plus a separate setting image if the selected workflow allows it. Do not substitute unrelated people and expect identity to merge correctly.

Picture 1 defines the character's face and outfit. Picture 2 shows the same
character from another angle. Picture 3 defines the room, not a new character.
The character enters the room, places one notebook on the table and stops.
Keep face, outfit and notebook color consistent. A fixed medium-wide camera
shows the full action. Sound: footsteps and the notebook touching wood;
no dialogue or music.

Replace these labels with the syntax your endpoint requires. This is a human-readable brief, not an executable JSON payload. Reference to Video explains how to assign roles and check continuity.

Revise one variable at a time

Symptom	First revision	What to inspect next
Face or product drift	Reduce movement; improve source clarity	Silhouette and fine details
Camera ignores the brief	Use one move and name its stopping composition	Framing and direction
Several actions collapse together	Split the brief into separate shots	Whether each clip completes its action
Audio is missing	Check the model’s audio support and controls	Actual output audio, not preview mute state
Unwanted music or speech	State its absence in the sound instructions	Listen to the full exported clip

Keep the same model and settings while testing a wording change. Record the model, prompt, references, duration, resolution, aspect ratio, seed if exposed, final quote and result. Only call a revision “better” after reviewing both outputs.

Budget for iteration

A successful but unsuitable clip still consumes Credits. Three successful attempts at the same settings cost about three times one attempt, before changing paid inputs. See the current cost examples; they are not official Flex prices. Pika’s hosted offering has its own pricing and feature details.

For official local setup and prompt examples, consult the model card. If you are unsure which model you are using, read the Flex versus H3 comparison first.