AI actor demo: sources

davidaistar videos (watched via analysis/gemini_.md + transcripts; paraphrased)

Rqim7SlMtZg — "TikTok Shop ads with AI actors" (motion swap)

• What he does, step by step: 1. Finds a TikTok Shop video with proven GMV (FastMoss/Kalodata). The example is an 11 s mirror-selfie shapewear clip lip-synced to a rap track, ~$11K GMV claimed. 2. Pulls the clean video and extracts the clearest full-body frame. 3. Swaps his AI actress's face and hair into that frame with an image model, keeping pose, light and clothes. The clothes stay because they ARE the product. 4. Runs a motion-swap video model (Kling Motion, no prompt) on original clip + new frame. 5. Optionally changes the voice or re-lips a new script.
• His own warning: swapping only the face over someone else's clip is close to copying a real creator's identity.
• Keep:
• The product stays REAL while the person is AI. This is the core of our demo rule; ours does it with the ugc-omni product gate (real pack refs, pick --product ≥7/10), with real footage as the fallback (Fish 10-06).
• Anchor on frame 1.
• Copy proven pacing, not invented poses.
• Change:
• No cloning other people's footage or identities: our motion/timing comes from OUR real b-roll and a real-creator reference frame (ugc-omni method, genetic change).
• Motion transfer / recast is not wired (registry: NOT BUILT).
• Lip-sync to music doesn't fit Dawn; we speak a callout.

gpzy2ofG0Dw — "duplicate viral TikTok Shop ads" (3 methods)

• Reference: a sleep-gummy "playful challenge" ad (couple asks "can these put us out in under an hour?", time-jump caption, one half-asleep whisper as proof), ~$68.8K GMV claimed.
• The three methods:
• (1) Product-swap only. He calls it grey-hat himself.
• (2) Full re-skin per ≤15 s segment, changing one thing at a time (actor first, voice in a separate pass for consistency).
• (3) His favourite: a vision model breaks the reference into per-second descriptions, an LLM chunks them into ≤15 s prompts, and the scenes are generated fresh.
• His fix for actors looking "too awake": name the exhaustion cues explicitly in the prompt.
• Keep:
• The challenge device (our register cell "challenge").
• One variable per edit pass.
• Name visible states in the prompt ("heavy eyelids"), not inferred.
• Method 3's dissect → chunk → prompt pattern. Ours is scripts/ad_dissector.py on non-competitor sources.
• Drop:
• Methods 1-2 on anyone's footage (10-05 rule; identity risk).
• The racial voice-swap prompt framing.

PmXqgiFn9_Y — "AI UGC & song ads for ~$1" (caveman talking head)

• Format: a 15-30 s realistic woman talking to camera in deliberately broken short English (the paraphrase shape is "noun bad, product good, now happy"). Script first.
• His voice journey: native video audio couldn't hold the exaggerated intonation, so he used TTS with stretched-vowel spelling plus lip-sync on 3-5 s audio slices. ~$1 and ~10 min per video claimed.
• Experimental and unconfirmed: cloning a real stranger's voice and ambience from a found clip.
• Keep:
• Short, start on the first syllable.
• A beat table (beat | time | line | visual).
• Cheap fast drafts, cull, upgrade survivors.
• Change:
• Our native-voice route already beats TTS for the actor (Fish 10-04). Stress comes from CAPS in the Omni dialogue, not letter-stretching.
• The caveman register is a gated cell (Fish's ruling).
• Never clone a stranger's voice: that's a likeness/consent problem, not a production trick.

CZwfTy7jZcQ — Seedance 2.0 vs Sora 2 hyper-real UGC

• Script shape: fast hook → problem → product → demo b-roll → CTA. That's this skill's beat order.
• His product-accuracy claim: omni-reference with many real product photos rendered gummy shape and a lipstick-peel demo correctly where other models failed. Extend-in-model keeps actor and voice across 15 s chunks.
• Keep:
• "Give it every real reference" as the TEST design for the product-fidelity isolation test (registry role: photoreal / multi-ref video).
• Extend as a candidate role (registry: video extend, NOT BUILT).
• Drop:
• The blurred / medium-distance face trick to get past likeness filters. Not adopted, ever.
• "Omni-ref solves product accuracy" as a blanket claim. Our 10-04 test with real hand-holding frames as refs still scored 3-4/10, so real footage stayed primary until ugc-omni's product gate (10-06: exact first frame checked vs the real pack, display-off wording, natural hold, locked product_action) superseded that rule.

diff_ai_ugc.md / diff_cloning.md (our diffs, 10-06)

• diff_ai_ugc top 5 says to pivot a test batch to his short 15-30 s native-voice format instead of the 80-90 s yapper story Fish closed on cost. This skill is that pivot, built on ugc-omni (one take per line) instead of the yapper keyframe chain.
• diff_cloning:
• The competitor-vs-organic source rule: dissecting any ad is fine; templating a lane leader's ad is not.
• The product-scale gotcha: never use a white-background product shot as the only scale reference. Use in-hand REAL photos.

Foreplay 06_yapper (Yuri collagen mask, 144 s AI talking head)

• Shows how a synthetic talking head plus picture-in-picture proof carries a sale.
• Tells to avoid: static car background, uncanny blinking, studio-clean TTS. ugc-omni's real-frame reference + native voice addresses all three.

Our receipts

• 10-04 yapper:
• Generated band-in-hand frames 3-4/10 (fused fingers, generic smartwatch look) even with real hand-holding b-roll frames as refs.
• Fish: "take out the dawn band with proper references". Product = real footage.
• Real b-roll cutaways from clip-library/raw/dawn-broll-2026-09-10/DAWN BROLL/Jessica/ worked as overlays.
• 10-04 23:05: Fish rejected ElevenLabs on the talking head and approved H3 native voice ("yea its good"). Native voice is the rule.
• 10-05:
• The band lying on the table, composited from the real cutout: product 7/10, reveal 5/10. The best actor-near-product option.
• Fish closed the yapper lane after the overbuilt keyframe chain. Simplest recipe first.
• 9-30 Side Effects (feedback_ai_avatar_broll_lane): motion proof first, one locked character, line-match ≥8/10, eyes over scores.
• Coded Dawn data (output/dawn-combo-map-2026-10-05/aggregate.txt):
• Static devices: callout-direct $190K ROAS 1.86; testimonial $26.8K 1.50; demo $153 1.10 (n1).
• cartoon-vo $314, ROAS 0 (n10).
• MAP §2B: in statics, a person reads as a testimonial, and testimonials lose. So the actor delivers a callout.
• Real footage inventory (scripts/clip_library.py, source dawn-raw, 89 clips registered 9-11 by scripts/register_dawn_broll.py): 66 tagged product, from creators Jessica, Kris and Lori. Wake-up, wrist, buzzing, fastening, out-the-door and talking-head shots exist.
• Real creator terms (ugc-engine/sops/04-outreach-sop.md, Fish 8-16): perpetual paid-ad usage, Meta only, $200/video cap, raw file required.

starpop.ai articles (his written process, read 10-06; paraphrased)

• how-to-use-seedance-2-0-to-make-ads / how-to-use-kling-3-0-to-make-ads:
• Changed: lock-take review order (identity → hands → line/sync/ending → continuity → camera → audio → caption room; correct beats pretty) and "diagnose the cause, change one thing" (§10). Clean-footage clause in rules (§6).
• Changed: reference map with one job per real product ref for the isolation test (§11).
• Confirms: one action per shot, shorten a line rather than speed it, put a hard-to-say term in an overlay, the generated actor is never presented as a real customer or expert (FTC note in the Kling article; TikTok requires an AI disclaimer).
• seedance-2-5-ads: its first proposed ad test (four product angles, one actor, one room, one motion ref; score product accuracy, hand interaction, rerolls, local-edit success) sharpened the isolation test. Its 30 s single-generation master is a registry candidate (Seedance 2.5 US), not a change here.
• hyper-realistic-ai-lip-sync-ads: the caveman park-bench ad was voice-first (TTS with stretched vowels) + a cheap fast lip-sync model on per-line audio slices, ~$1. Not adopted: native voice stays the rule for the actor (Fish 10-04); stress via CAPS, no letter-stretching; never clone a stranger's voice. Kept: per-line slicing so one bad shot is regenerated alone (we already run one Omni take per line).
• how-to-make-ai-ugc-videos-seedance-2-0 / seedance-2-vs-sora-2-ugc: "the actor holds and applies the product, hand movements nailed" and many-ref unboxing. Not adopted for the band: our generated band-in-hand is 3-4/10; superseded 10-06: generated product shots pass ugc-omni's product gate or fall back to real footage. Blur-face trick: not adopted (filter bypass). Green-screen UGC (actor over a study graphic) is a possible later register cell for the science/guarantee authority, but it needs a keying/composite step we haven't built; not in v1.