Pipeline Hub

format-motion-transfer

Motion transfer / actor swap: take only the MOTION of a source clip (choreography, camera move, comedic timing) and re-perform it with OUR cast, OUR set and OUR audio, either as a trend-gag hook spliced onto a paying Dawn body or as a new-cast fatigue refresh of our own winner.

Lane: motion-transfer/recast role (NOT BUILT: one-clip proof first) feeding cartoon-h3; photoreal people (target images, spoken lines, talking-head refreshes) = ugc-omni with its G0-G4 gates · not built yet: 6 (see below) · source: /Users/ayden/.openclaw/workspace/skills/format-motion-transfer

SKILL.mdbeat-sheet.mdsources.mdEvals (7)Learnings (11)
idpromptmustmust not
mt-01There's a viral TikTok of a girl doing a dance in her room. Swap her for our mom character and use it as the hook.runs the provenance gate first (source type, trend proof of >=5 unrelated accounts, no single creator's routine)
motion-control route only (generate from our image), never recast/edit on third-party footage
our cast from our sheets AND our set replace hers
source audio muted (keep_original_sound false + ffmpeg -an)
hook must show the morning problem; overlay line 1 names HER symptom
states the route is NOT BUILT and needs the one-clip proof first
i2i of our character into her frame
keeping her background or song
launching before the proof passes
a real person's likeness in the output
mt-02Take the HiStrips ad that's scaling and swap in our actor and our band.classifies HiStrips as a lane-leader competitor
refuses it as a motion source
routes to format-ad-clone as a gap map only
any model call on the competitor footage
product swap on their ad
keeping their script
mt-03Our UGC winner is fatiguing. Recast it with a new mom, same everything else.mode R with ONE variable = cast
confirms the footage is OUR own (not a charswap output derived from third-party footage) and the person in it consented
recast route (5-30 s per call, no shot >15 s, split on shot boundaries)
keeps our own audio; voice swap only as a separate pass
own ad set vs the parent, read CPA + frequency at 7 days
changing cast and voice in the same ad
judging before 7 days
pausing the parent
mt-04Run the motion transfer proof test so we can wire the lane.one clip only: Fish's own 5 s phone clip as the source
reads the fal schema + price page before quoting
P1 recast 768P ~$1.50 and/or P2 Kling motion-control with character_orientation video and keep_original_sound false, each <= ~$2 and each with Fish's go
pass criteria: beats within +-0.3 s, identity >=8/10 every second, no morph, no carry-over, Fish's eye decides
on pass: wire the entry point, update the registry row, append learnings
batch or wave generation
claiming the lane works before Fish sees it
judging by a metric alone
mt-05Make the transferred hook show the kid checking the band on his wrist when it buzzes.band stays out of the hook or wrists low and display dark
product beats come from the parent body under the wrist rule
drop sources where the arm rises to camera/face
raised wrist looking at the band
lit display
smartwatch look
mt-06Use the Kling motion thing we already have in the pipeline.says kling_motion in scene_parser/test_pipeline_fixes is only a prompt-text hint, not motion transfer
names the registry role 'Motion transfer / recast' and its untested candidates
lists what is NOT BUILT
inventing a script, flag or entry point that doesn't exist
hard-coding a model without naming its role
mt-07Write the motion-control prompt for the blanket-yank trend hook.names the ONE thing borrowed from the clip (the gesture + upper-body timing)
states face, hair, outfit and room come from our target image
source segment is one uninterrupted take
keep_original_sound set false explicitly
a long cinematic prompt restating the source's set or styling
a multi-cut montage as the motion reference
the source creator's audio