Mission
Produce a consented presenter or UGC-style draft from an authorized avatar image/video, script, and audio source. Keep visual identity, voice identity, spoken words, disclosure, and delivery format explicit. This skill does not authorize impersonation or publish synthetic media.
Consent gate
Before processing a real person's likeness or voice, confirm authorization for this production and intended distribution. Do not create non-consensual impersonation, sexualized deepfakes, deceptive political speech, or a minor's likeness. Record whether consent is unknown, claimed, verified, or approved for publication. A public profile or supplied file is not consent.
Collect:
- avatar/person ID and authorized image/video reference;
- script, pronunciation, language, captions, CTA, and disclosure text;
- audio source: supplied recording, approved generated speech, or silence;
- presenter framing, gesture, background, duration, ratio, and platform;
- rights for avatar, voice, music, logos, and location; and
- review owner, variant count, and cost/approval boundary.
Choose the path
| Inputs | Preferred path | Notes |
|---|---|---|
| Existing presenter video + authorized audio | media.lipsync |
preserve the parent video and inspect mouth/face sync |
| Authorized still/avatar image + audio | LTX or another live image/audio lipsync schema | use the current model schema; do not assume every lipsync tool accepts images |
| Image/video plus action or performance brief | media.generate_video |
use I2V/reference input, then inspect identity drift |
| No voice audio yet | adjacent voice agent or available speech capability | voice authorization and script approval still required |
Use MODELS.md and a live schema check. The current catalog contains more than one lipsync family and they accept different required URLs. The generic lipsync contract requires video and audio URLs; LTX 2.3's lipsync schema requires audio and may accept an image. Resolve exact fields through muapi-video-tools.md.
Execute
- Freeze the approved script. Mark every line as supplied, approved, or proposed; never add a factual claim for fluency.
- Upload local image, video, and audio inputs. Record roles, rights state, returned URLs, and expiry expectation without storing raw secrets.
- Check duration and segmentation. Split long speech into intentional segments only when the selected provider supports a reliable assembly path; preserve words and order across segments.
- Submit the selected job, retain its prediction ID, and poll or receive the result. Never silently switch a lip-sync job to a talking-avatar generator.
- Review mouth timing, phonemes, eyes, facial boundaries, head/hand motion, identity, background, captions, audio artifacts, and synthetic disclosure.
- Return the draft with a clear disclosure recommendation when a real person or generated voice is involved.
Use video QA and preserve parent/child links. If audio and video cannot be synchronized acceptably, report the failure and offer a new take or conventional edit rather than stretching speech silently.
Output contract
Return script/version, avatar and audio roles, consent state, model/endpoint,
prediction ID, output URL, status, duration, caption/disclosure treatment,
audio provenance, QA findings, and failed segments. Store a receipt under
.video/ only when permitted using
reports-and-receipts.md.
Publishing, creator outreach, and platform disclosure submission are outside this skill.