skills/pixverse-explainer/references/branches.md

# Specialized Faceless Branches

## Picture Story / Frame By Frame

One continuous natural narration → actual word timestamps → dense numbered frame timeline.
No video generation, per-phrase TTS, fixed 10s voice windows or slideshow substitution.
Target 0.7–1.2s/image, maximum 1.5s: at least ceil(actual_voice_seconds/1.5) frames, typically
45–70/minute. Show the planned image count before spending. Adjust an editable script from
measured speech to meet duration; never change audio speed.

A moment is a burst of 2–3 near-identical images changing one detail. Roughly two-thirds
of frames are literal edits of the immediately previous rendered frame as their ONLY
reference; preserve crop/camera/cast/colors/background and change one named detail. New
framing/action/place uses location→cast→props. Do not attach the roster to micro-edits,
which would redraw the whole scene. Keep each variation dependency sequential.

Mix wide/medium/close, show what speech names, and avoid consecutive close shots. The
source's close-up micro-variation wording conflicts with its no-two-close-frames rule;
resolve with one close reaction followed by a medium movement burst, keeping chains on
wide/medium shots. Frames have the output aspect, no roster sheets. Measured contiguous
segments have no gap/overlap and no hold >1.5s. Preserve continuous voice and deliver one film.

## Kids Narration

Block 1 asks the central question and answers it in ordinary words in the first sentence;
block 2 gives one visible example. Middle: what→where→does→without. Last: sourced surprise
then four/five-word callback answer. Read the first sentence of every block alone: they
must answer the opening question. One idea/sentence, at most two/block, familiar physical
analogies, no stacked abstraction, needless decimals or dates. Explain action before terms.

Every block changes the physical through-line. Narrator addresses named cast AND viewer;
cast visibly waves/nods/looks/responds on the beat, mouths closed. Later participation
questions may close a block, with the next block opening on their answer; this does not
delay the opening central question's same-sentence answer. Do not leave a long pause
inside either voice take. Warmth comes
from words and energy, not stretched syllables. Plan 17–21 English words/full 10s.

Four shots: context, character reaction, extreme detail and medium resolution, with order
and opening coverage varied between blocks. One real action spans them. Add 2–4 playful
diegetic cues tied to motion, no generated voice/music in narration clips. Kids looks
normally include a wordless bed, even on another channel using that look, with mood matched
to its content. Supplied music wins; use a supported music route otherwise. Start quietly
(~0.05 linear gain), duck and listen; gain is not a measured loudness claim. Explicit no-music
wins; disclose an unavailable optional bed while delivering the completed film.

## Kids Talking Characters

Opt-in. Odd blocks have external narration/closed mouths; even blocks have native dialogue
from at most two named speakers, 16–22 English words total/full10s. Dialogue is caused by
the preceding narration and sets up the next event. Keep identity/wardrobe/style/delivery.
Native timbre is not guaranteed identical across jobs; check if exact voice is required.
Keep dialogue's full-length audio aligned to lips; do not center or replace it with TTS.
Record block_kind and ordered speaker lines. Narration keeps its normal gate, all blocks
keep four-shot/music rules. Song intent selects Song instead of combining these modes.

## Kids Song

The primary audio must be a real sung song. Resolve genuine singing/music capability before
spending; TTS/instrumental music is not equivalent. Ignore narrator voice metadata, skip
narration takes and a second bed. Write visual lyrics in the requested language, verses
plus a chorus repeated verbatim. Generate song FIRST, then measure sections and choreograph.
Source targets 60/120s with ±3s tolerance and at most three attempts; explicit duration wins.

Gentle: 100 BPM major-key ukulele/guitar/bells/light flute. Dance: 112 BPM major-key claps,
plucked bass/marimba/bells. Keep steady 4/4 from start to end, no tempo changes or section
gaps; dynamics grow through instrumentation. Use regular singable lines, often eight
syllables/alternating stresses. Resolve the source's inconsistent eight-syllable/one-per-
beat/four-bar instruction: eight beats means two 4/4 bars; longer phrases need held notes
or rests. Specify an internally consistent meter instead of copying contradictory math.

60s: intro/verse, chorus, verse, fuller chorus. 120s: intro/verse 1, chorus 1, verse 2, chorus 2,
verse 3, final chorus. Verses advance story/show named nouns; choruses repeat signature dance
and location with growing cast/staging. Gesture/dance on beat, no articulated singing mouths.
Four-shot blocks, diegetic cues only; song owns the final audio, SFX barely audible (~0.06
starting gain). No automatic singing subtitles. Requested lyric captions need verified
singing alignment or actual manual timing, not an untested Whisper assumption.

## Long History (About 10 min+)

Outline chapters, eras, through-line, vignettes and asset counts first. Source skeleton:
two-block crisis opening, rewind around block 3, then 5–8 chronological chapters of 8–14 blocks
where runtime fits. Each chapter has setup→pressure→turn and a handoff hook, one documented
human vignette, a map beat and quantity comparison when figures are spoken. End with
meaning/consequence and one brief optional CTA. Adapt counts to actual runtime explicitly.

Verify 4–6 concretes/chapter and state historical uncertainty. One cast asset/era with fixed
identity invariants plus actual aging/clothing changes. Every later incarnation references
the first; for gradual 3+ eras also use the preceding era if needed. Match era variants to
blocks. Locations need variants only when changed. Sum cast×eras+places+coverage+props
before spending (often 25–40 assets/10 min, not a fixed quote).

Rotate scene→map→quantity→vignette; never three identical device classes in succession.
Action motifs occur at most once/chapter and about three/film, except through-line states
which must advance. Don't use seals/stamps as generic filler for named real content.
Produce chapter batches and preserve completed chapters; assemble one whole film. Mannequin
reduces eras to at most small/adult variants with stable role colors, no facial/costume aging.