THE photoreal people engine (Kristian Jennings method): real creator frame -> one-shot avatar (GPT Image 2, Nano Banana Pro split arm) -> locked Gemini Omni prompt -> A-roll per line chained on exact first/last frames (Omni 1.1 Flash) -> b-roll + product demo checked against the permanent Dawn ref sheet -> captions from the script. Two modes: cuts (A-roll + b-roll + scene change) and oneshot (one continuous yapper take). Format skills (ai-actor-demo, podcast B, street-interview, reaction-duet B, whistleblower, warehouse W3) hand it a job.json. Use for AI UGC, AI creator / talking-head ads, 'make it look like real UGC', or when an AI person has to show or demo the product. Not for cartoon (cartoon-h3).

Lane: — · not built yet: 1 (see below) · source: /Users/ayden/.openclaw/workspace/skills/ugc-omni

ugc-omni add-on — davidaistar AI-UGC teardown (2026-10-06)

Read-only add-on written by the format-skills session; ugc-omni's SKILL.md is owned by its own session and was NOT edited. Full step table with file:line citations: ~/research/davidaistar/analysis/diff_ai_ugc.md. Related format skills that use this lane: format-street-interview, format-podcast (photoreal execution), format-whistleblower (photoreal execution).

What he does that we haven't tested

Real Seedance 2.0. We've never run it. fal gives Seedance 1.5 Pro single-ref (scripts/seedance_pipeline.py, seedance_img2vid.py); the Higgsfield Seedance 2.5 unlimited entitlement ended 9-09. His results rest on 2.0's omni-reference (many inputs) and in-model extend. Cheapest access he names: StarPop trial or a ~$11/mo CapCut/Jianying plan. Needs Fish's go (money + new account). Best use: the Gate-1 motion proof that feedback_ai_avatar_broll_lane rule 1 requires. Hailuo i2v only held 1-3 s of motion, and that caused the Side Effects failure.
Omni-reference with many real product photos (front, side, in-hand for scale, packaging) instead of one ref. Our 10-04 hand-holding-band frames scored 3-4/10 (fused fingers, generic shape) on H3. One isolated test (the same band-in-hand shot, 6+ real photos, Seedance 2.0) tells us whether that is an H3 weakness or a real limit. Until it passes, real-footage compositing of the product stays the rule.
In-model extend for long takes, instead of our hand-built segment chain (scripts/yapper/yapper.py chain_plan/keyframes/render). Only worth testing once item 1 is live.
One variable per edit pass when re-cutting a take: actor, then voice, then product, never together.
Short native-voice formats (15-30 s) at ~$1-3 per test, instead of the 80-90 s single-narrator story that Fish closed on cost on 10-05. Fits the hook taxonomy better (starts in action, zero backstory).
Brand-DNA intake (product URL → images/colours/copy) instead of hand-assembling assets per batch. Not built.

Don't adopt

Getting around a provider's face or likeness filter (blurring faces, medium-distance tricks). Not adopted: compliance. Where reference-to-video refuses a realistic face, use i2v from a real-photo frame (ugc-omni's method).
China-proxy accounts, copyright-keyword bypasses, grey-hat product-swap-only clones of other people's footage.
"Omni-ref solves product accuracy" as a blanket claim. Our real-footage compositing stays primary for a small wearable until test 2 says otherwise.