Engine decision after this research (2026-09-28): Gemini Omni is the only video model. Where this report routes shots to Veo 3.1, the engine uses Omni with the rules on the main page.

Styles 02 · AI UGC and creator-native ad formats

Research date: 2026-09-27. Scope: user-generated-content (UGC) look and creator-native formats for 9:16, 30-60 s Meta / TikTok ads in Tunisian Derja (solar PV, alarms/cameras, then other SMEs). Research only. No media was generated.

Evidence tags: [official] = Google / Meta / TikTok / FTC docs. [vendor] = a tool maker's own claim. [practitioner] = agency or creator write-ups and tests. [secondary] = blog summaries of policy (check against the official page before relying on them). Most performance numbers in this field come from vendors and agencies. Read them as direction, not as proof.

Related docs: research/01-omni-prompting.md (Omni has no audio input; Derja VO unsupported), research/05-omni-pro-workflows.md, research/07-anti-ai-look.md (sound and finishing), research/08-hybrid-pipeline.md (Remotion RTL captions).

0. TL;DR

1. The script and the first 1.5-3 s decide most of the result, human or AI. UGC-look ads beat polished ads on hook rate and CTR in most agency data. For solar, agencies report that owner-on-camera and phone-filmed customer creative gives 25-40% lower CPL than stock or branded work [practitioner].
2. Derja lip-sync is the hard part. Veo 3.1 and Gemini Omni docs both say: "English (EN) is fully supported, but other languages have not been evaluated" [official]. Omni takes no audio input, so it cannot lip-sync to our Gemini TTS track [official]. Veo tends to pull the accent toward how the face "looks" [practitioner]. A native Derja talking head made only on Vertex is not reliable today.
3. There are two ways to get a talking head that speaks our Derja track. (a) Make the clip on Vertex (NB2 frame → Omni/Veo, "person speaking naturally"), mute it, then re-sync the lips to our TTS audio with a language-agnostic lip-sync model (sync. labs lipsync-2-pro / sync-3; sync-3 lists Arabic). (b) Use an audio-driven avatar (image + our audio): Kling Avatar 2.0, OmniHuman 1.5, Creatify Aurora, HeyGen Avatar IV/V, Captions Mirage. None of these runs on Vertex. Open models (LatentSync, InfiniteTalk, Wan2.2-S2V) can be self-hosted on a GPU. Google's new Gemini 3.8 Live Avatar (GA 2026-09-25, 97 languages) is conversational and uses its own voice. Custom avatars need allowlisting. It is not a fit for scripted ads yet.
4. Half of the formats need no on-screen lip-sync. Hands-only demo, POV, day-in-the-life with VO, green-screen with a STEG bill, review montage, split-screen and chat screens all work with our current engine. Ship these first. Add talking heads once a lip-sync A/B on Derja passes a native listener.
5. Policy decides how to frame it, not whether to use AI. A photorealistic AI person on Meta puts the "AI info" label next to "Sponsored" [official]. TikTok requires the AIGC label or a burned-in disclaimer. Undisclosed AIGC is rejected [official]. Fake customer testimonials are out on every platform and under Tunisian consumer law. Use an AI actor as spokesperson, a clearly labelled dramatization ("تمثيل / Mise en scène"), the owner's own avatar with written consent, or real customers filmed on a phone. For solar trust, the last option is still the strongest.

1. Why UGC and creator-native formats convert (data)

TikTok's own guidance [official]: 9:16, ≥720p, sound on, hook in the first 6 s and value in the first 3 s, feature people (creators, employees, customers), a DIY / not-too-polished style, text at 5-10 words per second, 3-5 creatives per ad group, and refresh when delivery declines.

What this means for Lead Up: the Tunisian audience is big and mobile-first. Facebook reaches 7.8M, TikTok 6.0M adults and Instagram 4.05M (DataReportal 2026). A cheap engine that makes many hook variants of the same proven body is the right fit, because UGC-style ads tire in about 2 weeks.

2. The core technical question: talking heads in Derja

2.1 What Vertex models can and cannot do (Sept 2026)

Conclusion: a Derja talking head made fully on Vertex is a gamble. It fails in the details natives catch: the voice drifts to MSA, Egyptian or Levantine, and it does not match our brand TTS voice. Keep our Gemini 3.1 TTS Derja track as the single source of truth and make the face follow it.

2.2 Lip-sync layers that take our audio (all off-Vertex)

Rules that hold across tools: audio first (expressive TTS gives expressive faces). Keep each take ≤8-10 s and cut, because drift and gesture loops show after that. Mid-shot, not extreme close-up. Use a face already in motion for redub models.

2.3 Test plan before we commit (no spend until the owner approves)

One 10 s Derja line with French code-switching, one NB2 first frame (Tunisian man, 35, car interior, phone selfie). Candidates: 1. Omni i2v "speaking naturally, no dialogue audio" → mute → sync-3. 2. Same → lipsync-2-pro. 3. Kling Avatar 2.0 (image + our WAV). 4. Veo 3.1 native, Derja in Arabic script inside quotes (control). Check the dialect by ear. 5. HeyGen Avatar IV (image + WAV).

Blind-rate with 3 Tunisians on: mouth match on emphatic consonants (ق ع ح ط), teeth and tongue, blink rate, "would you scroll past this as an ad or as a real video". The winner becomes the talking-head node in the pipeline.

3. Universal realism kit (applies to every format)

3.1 Realism killers → fixes

3.2 Caption style natives use

• TikTok/CapCut default look: bold rounded sans, white fill with black stroke or a white/black rounded box, 1-4 words per chunk, active word highlighted (yellow/green), lower-middle placement above the bottom UI safe zone (keep clear ~bottom 20% and right ~15% for buttons).
• Hook text as a separate box at the top third, 3-7 words, on screen from frame 1 (for sound-off viewers).
• Script choice for Tunisia: young Tunisians often write Derja in Latin "Arabizi" (3=ع, 7=ح, 9=ق: "9adech", "3aychek", "barcha"). Arabic-script Derja reads more "official". A/B both; for 35+ homeowners keep Arabic script plus French terms in Latin ("facture STEG", "garantie 25 ans"). Use Remotion RTL rules from doc 08 (per-word spans, FSI/PDI isolates for "300 DT").
• Native flourishes that help: comment-reply bubble, "Replying to @..." sticker look, emoji sparingly (☀️💡), a hand-drawn arrow or circle on proof.

3.3 Tunisian authenticity cues (settings, people, phone look)

• Settings: car interior (older Clio, Polo, Symbol or Picanto, sun visor down, phone in a dash mount); rooftop (سطح with white water tank, satellite dishes, solar water heater, neighbours' roofs, washing line); shop counter (حانوت, drinks fridge, scale, receipt book); home kitchen with tiled wall (زليج) and a gas bottle; café terrace (plastic or metal chairs, small coffee cups, "direct", tea glasses); villa gate and garden wall (alarm/camera); the paper STEG bill on a table. Avoid tourist clichés (blue doors of Sidi Bou Said) unless relevant.
• Summer heat and the "clim" (split AC unit on the wall) is the cause of bill shock in July-August. Show it.
• People: real mixes. Hoodie, bomber jacket or polo; installer in a hi-vis vest, cap and gloves; women with and without hijab; older man in a kachabia in winter scenes. Natural skin with shine and pores, a little sweat in summer, untidy hair.
• Phone look: mid-range Android front camera (Galaxy A / Oppo / Infinix style). Slight beauty smoothing, clipped windows, 30 fps, auto-exposure pumping when moving indoor→outdoor, vertical 9:16, sometimes a finger at the lens edge.
• Language: Derja base, with French for technical and money words (facture, installation, les panneaux, garantie, crédit, rendez-vous, franchement, normal). Money in دينار / DT / "mille" (spoken "280 alf" = 280 DT in some regions). Mention PROSOL Elec / STEG / ANME and bill-repayment credit only if the installer is actually eligible (subsidy ~500 DT, or ~1,500 DT for the economic variant, credit repaid on the STEG bill) [secondary; verify with ANME].

All Derja lines below are drafts and must pass the InfoLab Derja lint (reference/infolab/) and owner approval.

4. The formats

Difficulty is for our stack: ★ = current engine as is, ★★ = needs the lip-sync node, ★★★ = multi-speaker lip-sync or hard continuity.

4.1 Talking-head testimonial selfie ★★

• What: one person, phone at arm's length, talking about the result.
• Why it converts: testimonial is a steady Meta winner (6.5% hit rate, spend ratio 1.0, Motion). In solar, phone-filmed homeowner testimonials beat TV-style ads, and bill-number reversals are the top hook [practitioner].
• Script (30-40 s): 0-2 s bold number hook ("facture 380 DT… ولات 40") → 2-8 s pain (clim, summer) → 8-20 s what changed and how easy (one visit, 2 days, credit on the bill) → 20-28 s a small doubt removed ("كنت خايف من الـ arnaque…") → 28-35 s soft CTA ("اكتب 'solaire' في الكومنتار ولا اضغط على الزر").
• AI build: NB2 selfie frame → Omni 720p 4 s "speaking naturally, handheld" ×6-8 takes → mute → lip-sync to TTS (§2). Cut every 2-3 s. B-roll inserts (roof, inverter screen, bill) are made by Omni without faces.
• Killers: a stare that never breaks, TTS announcer tone, a spotless car. Fixes: eyes to screen, fillers, seatbelt on, parking-lot noise.
• Tunisia: a car in a parking lot at dusk, or the rooftop next to the panels.
• Compliance: an AI face presented as a real customer = fake testimonial. Not allowed. Variants: (a) real customer filmed on a phone (best); (b) AI actor, framed as a spokesperson ("أنا نحكيلك كيفاش يخدم…", with no "I bought") plus AI label; (c) "Dramatization based on a real client (Sfax, 2026)" on screen, and only with that real client's written consent and real numbers.

4.2 POV / storytime ★

• What: "POV: you just opened your July STEG bill". Drops the viewer into a scene, first person. It is the most shared TikTok format because people share "this is so me" [practitioner]. Motion lists POV among winning formats.
• Script: hook in <1.5 s (hooks under 2 s had 23% higher completion [practitioner]) → in-media-res scene → twist / solution → proof → CTA. Storytime = VO narrating over the scene ("خويا عامل الصيف الكل بالكليم…").
• AI build: no lip-sync needed. First-person shots (hands opening the bill, AC remote, rooftop from the eye line) from NB2 → Omni. Derja VO on top. This is our current engine exactly.
• Killers: cinematic grading, floaty camera. Fix: a phone-POV prompt with hand in frame, jerky pans.
• Compliance: fiction is fine, and clearly a scenario. The label comes from AI info. Keep claims real.
• Verdict: priority 1.

4.3 Street interview / vox pop (micro-trottoir) ★★★

• What: interviewer with a mic cube asks passers-by "قدّاش تخلّص في الـ facture STEG؟". Real reactions sell [practitioner]. Veo 3 AI street interviews went viral in 2025 because many viewers could not tell they were AI.
• Script: fast rhythm, 4-6 answers of 2-4 s each, one "surprised" answer, then the reveal ("وإنت، عملت الـ simulation؟"). 30-45 s.
• AI build: the hardest format. It needs many faces, two speakers per shot and lip-sync for each. Veo multi-speaker sync is below 40% even in English. Path: a single-speaker close shot per answer (the interviewer only as a mic hand and a VO) → lip-sync each answer. Street ambience is critical.
• Compliance risk: very high. A fake vox pop is a fake testimonial or public opinion. Only as an obvious sketch (named characters, humour, "تمثيل" label) or with real passers-by filmed by the installer. For opinion-style claims ("80% of Tunisians…") use real data only.
• Verdict: skip in AI. Offer as a real shoot service.

4.4 Podcast clip / two-person conversation ★★★ (★★ if one speaker)

• What: two people with visible mics (SM7B-style), 9:16 crop, one asks and the "expert" answers. It reads as a podcast moment, not an ad. 40%+ hook rates [vendor]. Script: organic hook → problem → product after trust. 30-60 s.
• AI build: a static two-shot of a café or studio set from NB2 (with a Tunisian "podcast" vibe: brick wall, neon, plants). Cut between singles (one face per shot), so each shot needs only one lip-sync. Listener reactions (nods, laughs) need no sync. Real podcast clips use 1-2 cameras and jump-cuts, which suits us.
• Script sample: Host: "ياخي الـ panneaux يخدمو في الشتاء؟" / Installer: "إيه يخدمو، أقل شويّة، أما الـ compensation متاع STEG تعدّل…"
• Killers: mics not in front of the mouth; audio too dry. Fix: "studio podcast" sound with close mic, slight compression and a tiny bleed of the other voice.
• Compliance: make the "expert" the real installer (owner avatar with consent) or an AI spokesperson labelled as such. Never a fake "engineer from STEG" (impersonating an authority is banned on TikTok and Meta).

4.5 Green-screen reaction ★ (★★ with face)

• What: the creator sits in a corner over a background screenshot (STEG bill, news headline on electricity price rises, Google Maps of the installer's jobs, a WhatsApp quote). You get proof and explanation at once [practitioner].
• AI build: the background is a real document or screenshot, redrawn cleanly in Remotion (never model-drawn). The foreground is a cut-out face. Option A: lip-synced avatar. Option B: faceless: a hand pointer and VO. Remotion does the chroma look, zooms and highlight boxes.
• Script: "شوفو هذي… (zoom on tariff tranche) هاذي علاش الـ facture طلعت" → explain → the fix → CTA.
• Compliance: use real articles and headlines with attribution. Do not fake a news outlet's page or logo (that is impersonation). Blur private bill data.
• Verdict: priority 1 in faceless form.

4.6 Unboxing / hands-only demo ★

• What: hands open the box (inverter, panels, alarm kit, camera), close-ups, a click-to-install moment, the app on a phone. Unboxing has the highest hit rate in Motion's Meta data (9.8%) and Demo is at 8.1%.
• AI build: NB2 frames with the real product photos as references (keep logos correct). Omni 4 s macro shots. ASMR sound (tape rip, foam, click). Derja VO or no VO. Hands are the main risk: keep takes short, cut before the fingers interlace.
• Script: 0-1 s sound hook (tape rip) → "شنوّة فيه الكيت متاع الكاميرا؟" → 3 features × 5 s → price / offer card → CTA.
• Compliance: the product shown must match what is sold (Meta and TikTok misrepresentation rules). Label AI.
• Verdict: priority 1 for alarms/cameras.

4.7 "Day in the life" of a tradesman ★ (★★ with speaking)

• What: from the 6 AM coffee to the rooftop install to the happy client, #dayinthelife / #bluecollar. It builds trust through behind-the-scenes (Motion lists Behind-the-scene as a steady winner). It fits the solar "anti-cowboy" trust play (a real team, a real van, real roofs).
• AI build: 8-12 shots of 2-3 s, mostly no-face or back-view (van, ladder, drill, panels, inverter LED, sunset roof), with a VO in the first person. Timestamps ("06:30", "11:00") are drawn in Remotion. One lip-synced line at the end is optional.
• Tunisia: coffee at the corner café, a loaded Isuzu/Peugeot pickup, the rooftop with a white water tank, lunch of kaskrout, the Maghreb call to prayer faint in the background (be respectful).
• Compliance: best done with real footage from the installer mixed with AI inserts. If it is all AI, label it and do not show a named real team member without consent.

4.8 Customer review montage ★

• What: fast cuts of review cards (Google Maps stars, Facebook comments, WhatsApp thank-you voice notes) with install photos. Social proof without faces.
• AI build: Remotion cards over Omni B-roll. The reviews must be real (screenshots with the customer's permission; blur names or use first names only).
• Compliance: never generate reviews. The FTC rule (16 CFR 465, in force since 2024-10-21) names AI-generated reviews explicitly. Tunisian law 92-117 art. 13 and 98-40 art. 35-36 ban misleading advertising.

4.9 Founder / owner-to-camera ★★

• What: the installer-owner speaks directly: origin story, "why we don't do cheap panels", a guarantee. Owner-on- camera = 25-40% lower CPL in solar [practitioner]. There are 11 founder sub-formats (Enemy, Origin, Underdog, Docu-style Q&A, Raw confession, Founder's call with a client, Silent proof…) [practitioner].
• AI build (best compliant AI use): a consented owner avatar. Use 20-60 s of real phone video of the owner plus a written release, then HeyGen twin / Kling Avatar / lip-sync on an Omni take generated from the owner's photo. The voice is our Gemini TTS or a consented clone. The owner approves every script. On Vertex, Live Avatar custom avatars need allowlisting. An even better hybrid: the owner films 5 real selfie clips once and we reuse them with new VO via lip-sync (a redub of real footage looks more real than any generated face).
• Script: "أنا [name]، عندي 12 سنة نركّب في [city]. اليوم باش نقلّك علاش ما نركّبش panneaux رخاص…" → one enemy → one proof → offer.
• Compliance: consent file stored. The AI label still applies (a photorealistic AI person). Never make the owner claim results he cannot prove.

4.10 Duet / stitch / reply-to-comment ★ (★★ with face)

• What: a native TikTok grammar. "Replying to @user: 9adech ta3mel l'installation?" with a comment bubble, then the answer. It uses real objections as hooks and feels like a conversation. Spark Ads keep duet/stitch features and lift completion [official-ish].
• AI build: a Remotion comment-bubble sticker (TikTok style) over a talking head or a faceless demo. Stitch = the first 2-3 s of a (real, licensed or our own) clip, then our answer.
• Compliance: the comment must be a real comment (from the page or from DMs). Anonymise the handle. Do not fabricate a user.
• Verdict: a great hook-variant generator. The same body can take 5 different real objections.

4.11 "Us vs them" / split-screen ★

• What: left = cheap installer (dusty panels, no guarantee, bill still high), right = us. Or "STEG only vs STEG + solar". Both sit in Motion's top hit-rate band.
• AI build: two Omni takes with matched framing, composited in Remotion with labels. No lip-sync needed.
• Compliance: comparisons must be factual and verifiable. No naming or showing an identifiable competitor unless the claim is provable. TikTok bans misleading before/after that "may cause viewers to have a false or distorted impression" [official]. Use a generic "installateur pas cher".

4.12 Other UGC-native formats that perform in 2026

5. Ethics and platform policy

5.1 What the rules say

• Meta [official]: the "AI info" label covers ad images and videos made or significantly edited with Meta's or third-party gen-AI tools (detected via industry signals such as C2PA). If the ad has a photorealistic AI person, the label sits next to "Sponsored". Otherwise it sits in the three-dot menu. Political/social ads need self-disclosure. Secondary sources say detection of third-party tools expanded from 2026-06-01, and Google outputs carry SynthID and C2PA, so assume every Veo/Omni ad gets labelled.
• TikTok [official ad policy]: "Significantly edited media and AIGC content is allowed if… Apply the AIGC label, or… a clear disclaimer, caption, watermark, or sticker." Undisclosed AIGC is "rejected or restricted". Public-figure likeness without permission is banned, and so are misleading before/after results. Secondary sources [unverified] add: consent paperwork for voice clones and likenesses in Ads Manager, the toggle resets when a campaign is duplicated, auto C2PA labels cannot be removed, virtual influencers may state facts but not personal endorsements, and TikTok Shop bans AI fake testimonials with a risk of permanent bans.
• FTC 16 CFR 465 (US, the global reference standard): bans fake or AI-generated reviews and testimonials, including from people who don't exist or have no experience with the product. Penalties up to ~$53k per violation. The FTC Q&A says "using actors to portray testimonialists could be deceptive" unless the dramatization is obviously fictional. There is no blanket ban on AI avatars [official].
• Tunisia: Law 92-117 (consumer protection, art. 13) and Law 98-40 (sales techniques and advertising, art. 35-36) make misleading advertising an offence once it is broadcast in Tunisia [secondary]. Personal-data law (Loi organique 2004-63) should cover using a real person's face or voice. Not verified in this pass. Get a local legal check before the first owner-avatar campaign.

5.2 Compliant variants (use these names in scripts and briefs)

House rules: 1. Always toggle the platform AI disclosure (Meta and TikTok) when a synthetic person appears. Also add a small burned-in "AI" or "تمثيل" corner tag for dramatizations. 2. Numbers (savings, prices, subsidy) must come from the installer's real jobs and be documented. 3. Owner approves every script before any paid generation (existing rule). 4. Never depict minors, public figures or identifiable private people. 5. Keep a per-ad compliance note (format, disclosure, consent file, claim sources).

6. Recommended AI UGC production recipe (our Vertex stack)

Brief + real proof (bill numbers, install photos, reviews, owner clips)
  → Derja script (InfoLab lint, owner approval)            [gate]
  → Gemini 3.1 TTS: casual Derja VO, fillers + breaths      (source of truth)
  → Timing spine (word timings → shot list, cuts every 1.5-3 s)
  → NB2 first frames: character sheet (3 refs), Tunisian set, phone-selfie look, real product refs
  → Omni 720p i2v, 4 s takes (always 720p), prompts: "front-camera phone video, handheld, flat light,
       [speaking naturally | no dialogue]"; B-roll/hands/POV with no faces talking
  → TALKING SHOTS ONLY: mute → lip-sync node (winner of §2.3 test: sync-3 / lipsync-2-pro / Kling Avatar)
       fed with the exact TTS slice; ≤8 s per take
  → Remotion: native captions (Arabic script or Arabizi A/B), hook box, comment bubbles, bill/review cards,
       split screens, AI/تمثيل tag, CTA end card
  → Sound: phone-mic EQ on VO, room tone per set, foley, light music bed (doc 07 §3)
  → Finish: grain, slight over-sharpen, 30 fps, -14 LUFS
  → ffmpeg export 1080x1920 → platform upload with AI disclosure toggled

Rollout order: - Wave 1 (current engine, no lip-sync): POV/storytime, hands-only unboxing/demo, day-in-the-life with VO, faceless green-screen bill, review montage (real reviews), split-screen, chat-screen, bill reveal. - Wave 2 (after the §2.3 lip-sync test): owner-to-camera avatar (with consent), AI spokesperson selfie, car "spokesperson", reply-to-comment with face. - Wave 3 (hard, low priority): podcast two-shot, solo skit. Street interview stays a real-shoot service only.

Variant economics: one body (20-30 s) × 5 hooks (bill number, POV, real comment, confession, contrast) × 2 caption scripts = 10 ads per concept. Refresh every 10-14 days.

7. Open questions

1. Does Veo 3.1 speak recognisable Tunisian Derja from an Arabic-script quote, and does the accent survive on a Tunisian-looking face? (One LinkedIn practitioner claims Veo 3 does "local dialects"; not verified.)
2. Is a lip-sync API allowed under our data rules (upload of VO and owner face to a US/Chinese vendor), or do we self-host LatentSync/InfiniteTalk (the Runpod plugin is present but needs authentication)?
3. Can we get Live Avatar custom-avatar allowlisting for the owner, and can it read a fixed script with an external voice?
4. Arabizi vs Arabic-script captions: which gets better hook rate on Tunisian TikTok vs Facebook?
5. Local legal check of 92-117 / 98-40 / 2004-63 for AI dramatizations and owner avatars.

Sources

Official - Gemini API, Omni Flash: https://ai.google.dev/gemini-api/docs/omni - Gemini API, Veo 3.1: https://ai.google.dev/gemini-api/docs/veo - Vertex Veo 3.1 model page: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/models/veo/3-1-generate - Gemini 3.8 Live + Live Avatar GA: https://cloud.google.com/blog/products/ai-machine-learning/gemini-3-8-live-with-live-avatar-is-now-generally-available - Live avatars docs: https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/live-api/configure-live-avatars - Google Vids avatars, 24 languages incl. Arabic (June 2026): https://workspaceupdates.googleblog.com/2026/06/enhanced-ai-avatars-vids.html - Google Vids languages (Feb 2026): https://workspaceupdates.googleblog.com/2026/02/vids-avatars-voiceovers-new-languages.html - Google Vids avatar directing (Apr 2026): https://workspaceupdates.googleblog.com/2026/04/direct-avatar-to-speak-and-act-anywhere-with-a-consistent-face-and-voice.html - Veo prompt guide: https://deepmind.google/models/veo/prompt-guide/ - Meta, GenAI transparency in ads: https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/ - Meta help, AI labels in ads: https://www.meta.com/help/artificial-intelligence/355108217670024/ - Meta business help, third-party AI info: https://www.facebook.com/business/help/1009525374368348 - Meta AI disclosures: https://transparency.meta.com/policies/other-policies/meta-AI-disclosures - TikTok ad policy, misleading and false content: https://ads.tiktok.com/help/article/tiktok-ads-policy-misleading-and-false-content - TikTok creative best practices: https://ads.tiktok.com/help/article/creative-best-practices - TikTok disclaimers in Ads Manager: https://ads.tiktok.com/help/article/add-disclaimers-to-ads - TikTok Spark Ads creative playbook: https://ads.tiktok.com/business/creativecenter/quicktok/online/Spark-Ads-Creative-Playbook/pc/en - FTC final rule on fake reviews: https://www.ftc.gov/news-events/news/press-releases/2024/08/federal-trade-commission-announces-final-rule-banning-fake-reviews-testimonials - FTC rule Q&A: https://www.ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule-questions-answers - eCFR 16 CFR 465: https://www.ecfr.gov/current/title-16/chapter-I/subchapter-D/part-465 - Tunisia Law 92-117 (Juridoc): https://www.juridoc.tn/document/loi-n--1992-117-du-7-decembre-1992-relative-a-la-protection-du-consommateur/26006?_locale=en

Tools / vendors - sync. labs lipsync models: https://sync.so/docs/models/lipsync ; pricing https://sync.so/pricing - Sync LipSync 3: https://wavespeed.ai/blog/posts/introducing-sync-lipsync-3-on-wavespeedai/ - Creatify Aurora: https://creatify.ai/introducing-aurora ; best practices https://creatify.ai/blog/aurora-best-practices - Captions Mirage: https://captions.ai/blog/mirage-worlds-first-foundation-model-for-ugc-video - HeyGen Avatar IV: https://www.heygen.com/avatars/avatar-iv ; review https://diyai.io/ai-tools/video-generation/reviews/heygen-review/ - Arcads languages: https://intercom.help/arcads/en/articles/13250377-what-languages-are-supported-by-arcads - Kling Avatar 2.0: https://www.therundown.ai/tools/kling-avatar-2-0 - OmniHuman 1.5: https://wavespeed.ai/models/bytedance/avatar-omni-human-1.5 - Wan2.2-S2V: https://huggingface.co/Wan-AI/Wan2.2-S2V-14B - Higgsfield lipsync guide: https://higgsfield.ai/blog/make-ai-lipsync-videos - Invideo lipsync model comparison (Aug 2026): https://invideo.io/blog/ai-lipsync-models/

Practitioner / analytics - Maxfusion, lip-sync test of models for AI UGC (2026): https://maxfusion.ai/blog/ai-video-models-lip-sync-2026 - Prompt Architects, dialogue and lip-sync by model: https://prompt-architects.com/blog/310-dialogue-and-lip-sync-in-ai-video - Veo non-English accent issue (Romanian) and redub workaround: https://github.com/ulmeanuadrian/kie-mcp/issues/1 - Veo 3 prompting guide (selfie, dialogue colon trick): https://github.com/snubroot/Veo-3-Prompting-Guide - Atlas Cloud, Gemini Omni hands-on: https://www.atlascloud.ai/blog/tips/google-gemini-omni-testing-results - VentureBeat, Gemini Omni launch: https://venturebeat.com/technology/google-unveils-gemini-omni-any-to-any-ai-model-what-enterprises-should-know - Motion Creative Benchmarks 2026, formats: https://motionapp.com/library/research/creative-benchmarks-2026/top-visual-formats - Motion benchmarks overview: https://motionapp.com/library/research/creative-benchmarks-2026/ - Foxwell on Motion benchmarks: https://www.foxwelldigital.com/blog/motion-creative-benchmarks-2026-8-key-takeaways - inBeat, AI UGC vs real UGC: https://inbeat.agency/blog/ai-ugc-ads-vs-real-ugc - Finsi, UGC benchmarks 2026: https://www.finsi.ai/blog/ugc-ads-performance-benchmarks/ - Benly, creative benchmarks 2026: https://benly.ai/learn/ad-creative/ad-creative-benchmarks-2026 - Digital Applied, AI creative benchmarks: https://www.digitalapplied.com/blog/ai-ad-creative-benchmark-2026-ctr-roas-data - Pose, AI UGC vs creators: https://pose.ai/blog/ai-ugc-vs-real-creator-performance - Z2A Digital, making AI UGC human: https://www.z2adigital.com/blog-content/make-ai-ugc-human - Foreplay, 10 ad formats: https://www.foreplay.co/post/savanah-sanchez-ad-formats - The Performers, founder ad playbook: https://www.blog.theperformers.io/p/the-founder-ad-examples-formats-that-convert - Influee, podcast-style ads: https://influee.co/landing/podcast-style-ads - NoGood, man-on-the-street interviews: https://nogood.io/blog/man-on-the-street-interviews/ - Veo 3 local-business ads: https://www.joeyoungblood.com/artificial-intelligence/local-businesses-are-creating-viral-videos-ads-with-googles-veo-3/ - TikTok hooks / POV: https://zeely.ai/blog/hooks-for-tiktok-video-ads/ ; https://influee.co/blog/tiktok-ad-examples - Spark Ads data: https://www.adsmurai.com/en/articles/tiktok-spark-ads ; https://reugc.com/blog/tiktok-spark-ads - Solar ads: https://leadsync.me/blog/solar-facebook-ads/ ; https://www.webtonic.io/blog/solar-energy-ad-creative-statistics

Secondary policy summaries (verify against official pages) - https://www.cinerads.com/blog/ai-ugc-facebook-ad-policy - https://www.cinerads.com/blog/tiktok-ai-content-policy - https://ugcvids.ai/blog/tiktok-ai-content-disclosure-rules-2026 - https://coinis.com/blog/meta-ai-content-labeling-facebook-instagram-ads-2026

Tunisia context - DataReportal Digital 2026 Tunisia: https://datareportal.com/reports/digital-2026-tunisia - NapoleonCat Tunisia (Jul 2026): https://stats.napoleoncat.com/social-media-users-in-tunisia/ - PROSOL Elec subsidies: https://www.stes.tn/subventions-photovoltaique-tunisie-prosol-anme/ ; https://tunisie-electricite.com/steg/solaire - Tunisian influencer marketing (Derja/French): https://dergipark.org.tr/en/pub/tujomr/article/1731207