Pika H3 Flex Prompts: Five Copyable Patterns for Video with Separated Audio

Pika H3 Flex is driven five ways — text-to-video, first and last frame, multimodal reference-to-video, editing an uploaded clip, and audio-stem export — and each way wants its prompt written differently. These five patterns are copyable: replace the parts in [brackets], keep the structure. Square-bracket placeholders stay in English whatever language you prompt in. Last checked against Pika's published limits: 2026-11-03.

Before you prompt: the limits that shape the prompt

• 5 to 15 seconds, whole seconds only, at 24 fps. Write the shot to fit; a prompt that needs 20 seconds will not get it.
• Prompts up to 8,000 characters. Long, specific prompts are legal; use the room.
• Up to 12 reference assets — as many as 9 images, 3 videos and 3 audio files. Video and audio references run 2–15 seconds each.
• Native resolution is 480p or 768p; 1080p, 2K and 4K arrive through Pika's in-job upsampler, finishing the same take.
• Audio is native: stereo, with dialogue, score, SFX, foley and ambience exportable as five separate stems.

1. Text-to-video with dialogue and sound design

Name the shot, the camera, the action beats in order, and then the sound — dialogue in quotation marks, each sound layer named. The model composes what you describe; what you leave out, it chooses.

"A 12-second 16:9 medium-wide shot at dusk: [a cyclist in a yellow rain jacket] coasts down a wet cobblestone street as the camera tracks alongside at handlebar height; she brakes at a bakery window, plants one foot, and says \"One loaf, the one with the seeds." A shop bell rings as the door opens behind her; light rain patters on her hood; a quiet piano score sits under everything."

Check: the line lands on the lip movement, the rain is steady rather than flickering, and the take is one continuous shot with no hidden cut.

2. Reference-driven character spot

In reference mode the assets carry identity and the prompt carries direction. Say what the references hold, then direct only what happens. Never re-describe the character's face in words — the reference owns that.

"Using the attached character reference for [the presenter] and the product reference for [the bottle]: a 10-second 9:16 shot in a bright kitchen. She lifts the bottle, reads the back label, nods once to camera, and says \"Actually — same price as last year." Camera holds still; morning window light from the left; kitchen ambience low."

Check: the face and clothing match the reference in the last second as closely as the first, and the label text survives being read on camera.

3. First and last frame

Give Pika the opening frame and the closing frame, and write the prompt as the path between them: what changes, in what order, at what pace. Frame inputs are images; the prompt is the verb.

"From the first frame to the last frame: a 8-second 21:9 shot where the tide comes in around [the beached wooden boat], each wave reaching further up the sand; by the final frame the hull floats. Gulls cross once mid-shot; wind and waves only, no music."

Check: the first and last rendered frames genuinely match the two images, and the progression is monotonic — the tide should not retreat mid-take.

4. Editing an uploaded clip

For edits, upload the source clip as the video reference and write the change as an instruction, then say what must not change. H3 Flex preserves performance and timing — the prompt's job is to bound the change.

"Edit the attached clip: replace [the gray hoodie] with [a navy peacoat], keep everything else identical — same walk, same speed, same camera, same street and weather. 10 seconds, audio unchanged."

Check: the gait and stride timing match the original frame for frame, and nothing else in the scene drifted — store fronts, passers-by, light.

5. Stems-first audio direction

When the audio matters more than the picture — dialogue mixes, localization, a score swap later — direct each stem by name. Exported stems arrive as five WAVs: dialogue, score, SFX, foley, ambience.

"A 10-second 16:9 shot of [a barista] working behind the counter at opening time, camera slowly pushing in. Dialogue: she says \"First one's on the house." Score: none. SFX: the grinder runs, the portafilter clicks in. Foley: her apron rustling, cup set on saucer. Ambience: quiet café room tone, a freezer hum in the right channel."

Check: each stem contains only what its line named — the grinder should not bleed into foley — and the dialogue stem is usable alone.

Running these prompts

H3 Flex runs through Pika's API or product (or self-hosted weights) — it is not on Imgo. The same briefs run today on Seedance 2.0 in Imgo's video generator, which takes image, video and audio references and generates native audio in the prompt, with every run quoting its credits before you submit. Rates and limits for both models are compared in the Pika H3 Flex pricing breakdown.

Sources: Pika's H3 Flex model page and launch post, checked 2026-11-03.