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 01 · Photoreal, hyperreal and cinematic ad styles (Vertex-only engine)

Research date: 2026-09-27. Scope: live-action-look ad styles made with AI for 9:16, 30-60 s Meta/TikTok ads in Tunisian Derja (solar PV, alarms/cameras, later other SMEs). Pipeline constraint: Google Vertex only (Nano Banana 2 = gemini-3.1-flash-image, Gemini Omni gemini-omni-1.1-flash-preview, Veo 3.0/3.1 incl. fast/lite, Lyria 002). Companion docs: ../07-anti-ai-look.md (generic finishing chain, sound), ../06-nano-banana-2.md (NB2 references), ../01-omni-prompting.md (Omni anatomy). This doc does not repeat those; it adds the photoreal-specific layer.

Tags: [official] = Google/Meta/TikTok docs; [practitioner] = guides, reviews, case studies; [ours] = our synthesis or estimate, not verified by a source. House rules still apply (Omni 720p, no last frame on Omni, all text in Remotion, script approved before any paid generation).

0. TL;DR

1. Route by realism risk, not by habit. Veo 3.1 Standard is still the photoreal reference for faces, skin, light and product fidelity; Omni Flash is cheaper per iteration and strong on object physics, but its faces drift on head turns, motion can feel floaty and can drop to a choppy 8-12 fps look, and lip sync degrades after ~6-7 s [practitioner: aireiter, invideo Omni review, buildfastwithai]. Rule [ours]: faces/hero product = Veo 3.1 (Fast for takes, Standard for finals); B-roll without faces, drones, hands-free inserts = Omni 720p.
2. Realism is decided in the first frame. Every photoreal shot starts as a Nano Banana 2 still (or Nano Banana Pro for close-up faces, if enabled on the project) with camera body, lens, light direction and imperfections named, then animated with a short motion-only prompt. Image-to-video beats text-to-video for realism [practitioner: Sunra, Runway, Hailuo].
3. For lead-gen, "native realism" converts better than "cinematic polish". Native, handheld, phone-look Reels beat polished commercial creative by 25-40% on engagement (vendor benchmark); for solar, phone-filmed testimonials and owner-on-camera cut CPL 25-40%, bill-shock hooks with real bills beat abstract savings 3-5x, before/after install videos give ~2x leads vs static [practitioner: digitalapplied, webtonic, leadsync]. Cinematic is for brand films, retargeting and premium SMEs; the lead ad is usually "shot on iPhone" + documentary + real bill.
4. Being perceived as AI costs money; the label does not. NIQ neuro-testing: AI ads read more "annoying, boring, confusing" with weaker memory activation; when users perceive AI, premium perception -17%, purchase intent -14%. But a 900-person MediaScience/Ehrenberg-Bass study (June 2026) found AI disclosure labels barely move recall, recognition or sentiment. So: always disclose, and spend the effort on not looking AI.
5. Compliance traps specific to us: (a) Veo personGeneration in MENA/EU is allow_adult only, and TikTok removes synthetic likenesses of minors: no realistic children, ever; (b) an AI person presented as a real customer ("Sami from Sfax saved 70%") is a fake testimonial on Meta/TikTok/FTC standards, even if labelled; use a spokesperson or dramatization framing, and real customers for testimonials; (c) Meta may show the AI label next to "Sponsored" for photoreal AI people, TikTok requires the AIGC toggle for realistic people/scenes.
6. Tunisian authenticity is a prompt + reference problem. Models default to Orientalist "exotic East" imagery (4.02 cultural symbols vs 0.20 political/modern per image for Eastern countries in a 2025 study). Counter it with "contemporary 2026 Tunisia", concrete everyday details (beige/white rendered villa, fer forgé grilles, flat roof with solar water heater and rebar, white louage with red stripe), and real location photos from the installer as NB2 references.

1. What we can use on Vertex (facts that shape every style)

Quality consensus (2026 reviews): Veo 3.1 "photoreal king" for faces, skin with pores, lighting that respects the scene, and the most predictable product consistency across generations; Seedance 2.0/Kling 3 lead on fast action physics and cloth/liquid (not available to us); Omni Flash "correctness over polish", strong object physics, weaker human motion and faces [practitioner: MindStudio, magicshot, wavespeed, aireiter, invideo].

Implication [ours]: the photoreal engine needs a second video route (Veo 3.1) next to Omni. Keep Omni for what it is good at; do not force faces through it.

2. The photoreal master workflow (what top AI studios do, adapted to us)

2.1 Pipeline

1. Script + shot list (approved). Each shot gets: sub-style, lens, light, one action, duration kept (1-3 s).
2. Casting (once per campaign). NB2/Pro portrait → 8-view character sheet (front, back, 2 profiles + 4 face close-ups) on neutral background at 2K-4K; generate 4-5 options, lock one; one documented production averaged 5 generations (~$10) per character [practitioner: invideo, apipass]. New costume = new sheet, never edit the old one. Remove held props from the sheet (props drift between angles).
3. Location plates. Real photos of the client's real site (roof, shop, villa) shot on the owner's phone, used as NB2 references; or NB2-generated plates locked like characters ("Gatsby mansion lobby created once, then referenced throughout" in Higgsfield's $350k-look perfume spot) [practitioner].
4. First frames. NB2 composites: character sheet + location plate + product → one 9:16 2K still per shot, lens and light named. Generate 3-4, pick 1. This is where realism is won.
5. Motion. Omni 720p (no faces, simple moves) or Veo 3.1 Fast i2v (faces, hero product), then re-render only the winners on Veo 3.1 Standard 1080p if needed. 3 takes, keep 1-3 s.
6. Conform + upscale. One fps, one resolution. Topaz Starlight Precise 2.5 (built to reduce "plastic" artifacts on AI footage without inventing content) for faces/products; Astra 2 only for wide, detail-poor shots (creative diffusion, invents detail) [practitioner: invideo Topaz]. Or Veo native 1080p/4K.
7. Grade + texture (section 4), sound (see 07 §3), Remotion text/UI, ffmpeg master.

2.2 Consistency levers (strongest first) [practitioner + ours]

• Same NB2 first frame lineage (composite the character into each scene before animating) > reference images alone.
• Veo: 3 asset refs max, so spend them on character + product + location, not 3 angles of a face. Frames and references are mutually exclusive in some APIs: when you pass a first frame, references may be ignored; bake identity into the first frame [practitioner: Vercel/liteLLM docs].
• Re-describe character, wardrobe, setting and light verbatim in every prompt (PJ Accetturo: "re-describe settings, characters and tone in each prompt") [practitioner].
• One light direction and colour temperature for the whole ad; lighting mismatch between cuts is a top AI tell ("inconsistent background physics between cuts") [practitioner: OpusClip].
• Veo seed for re-rolls of a good take (not deterministic). Omni has no seed: use its conversational edit on the winning clip instead of regenerating.
• 2-3 characters max per scene; consistency falls apart at 4+ [practitioner: apipass].

2.3 Hit rates, time, cost

• PJ Accetturo's Kalshi NBA Finals spot (Veo 3, text-to-video, 2025): 300-400 generations for 15 usable clips (~5%), one person, 2-3 days, ~$2,000 AI cost, "95% cheaper" than a shoot [practitioner]. Liquid Death spec: ~$800, 2 weeks.
• Image-first i2v raises the keep rate a lot; our planning estimate is 1 keeper per 3 takes on Omni/Veo Fast with approved first frames [ours, to be measured].

Cost per 30 s photoreal ad [ours, list prices, 14 shots, 3 takes each, 4 first-frame candidates each]:

Time [ours]: 1 working day of generation + 0.5 day finishing for route B once the style kit exists; the first ad in a new style costs 2-3 days (casting, location plates, grade LUT).

3. Sub-styles

Format for each: look · why it converts · best model (market) → how we get it on Vertex · prompts · realism killers · post · Tunisian fit. Prompts are English (models are English-first; Omni/Veo "other languages untested") [official]. Motion prompts are motion-only for Omni (house rule); for Veo add "no music, no dialogue" because we replace audio.

3.1 Hyperreal product / B-roll macro ("commercial tabletop")

Look. Probe-lens and 100 mm macro, extreme close-up that reveals the product slowly; hard single key or big diffused panel, rim light, black or stone/travertine sweep; liquids with caustics, condensation, fresnel on glass, anisotropic highlights on brushed metal; motorised slider/dolly or turntable; 24 fps or "high-speed" look; grade clean, neutral-warm, deep blacks.

Why it converts. Texture-first macro is the thumb-stop for product categories; Veo 3.1 is strongest on texture-rich items, liquids/refraction, metal and minimal packaging [practitioner: picassoia]. For us: solar cell texture, inverter LEDs, MC4 connectors clicking, rain beading on a panel, a camera lens iris, an alarm keypad.

Model. Market: Veo 3.1 (product consistency), Seedance 2.0 (liquid physics). Ours: NB2 first frame → Veo 3.1 Fast i2v, final on Veo Standard 1080p for liquids/glass; Omni fine for rigid objects and turntables.

Image prompt (NB2, 9:16, 2K).

Extreme macro product photograph of a monocrystalline solar panel corner, black cells with fine silver busbars, anodised aluminium frame, fresh rain droplets beading on the anti-reflective glass. Shot on a 100mm macro lens at f/5.6, key light a large diffused softbox from top-left, thin warm rim light from behind, black seamless background. Commercial tabletop photography, true material texture, subtle dust specks on the frame, no text, no logos.

Motion prompt (Veo/Omni).

Single continuous shot. Slow motorised slider move left to right, 2 cm of travel. One droplet slides down the glass and merges with another, surface tension visible. Panel shape, frame and cell grid stay identical, no deformation, no flicker. Locked focus plane.

Realism killers → fixes. Melting/deforming product (add a consistency guard sentence like the one above [practitioner: awesome-ad-video-prompts]); "extreme shine" plastic metal (ask for "soft glow alongside specular", describe the light source before the highlight, "reflecting a studio environment") [practitioner: Hailuo]; flickering highlights (one key light only); gibberish logos/labels (text-free generation, logo composited in Remotion); too-clean surfaces (add dust, fingerprints, micro-scratches).

Post. No heavy grain on macro (1-2%), slight highlight halation only for luxury variants, true blacks, 2-3% vignette. A 110-130% punch-in on a 1080p plate gives a second shot for free.

Tunisia. Tunisian sun is the product: real Mediterranean hard light, fine orange Saharan dust on the glass (chergui/sirocco days), olive leaf fallen on the panel for Sfax, sea-salt haze for Sousse/Monastir/Djerba.

3.2 Cinematic brand film / narrative short

Look. Anamorphic character (oval bokeh, horizontal flares, slight edge distortion), 35-50 mm for people, 85 mm for intimacy, 16-24 mm for space; motivated light (practicals, window, golden/blue hour); film stock emulation (Kodak Vision3 500T for night/tungsten with halation, 250D daylight, Portra-like warm skin), 24 fps with 180° shutter blur, restrained grade (not teal-orange cliché). Story in 3-5 beats: problem → turn → relief.

Why it converts. Builds brand/trust for higher-ticket installs and is the premium differentiator for Lead Up's agency pitch; weaker than native formats for direct response. Hybrid work (AI + human craft finishing) avoids the backlash that purely generated spots triggered (Coca-Cola, Progressive negative sentiment; Kalshi and Liquid Death worked because they leaned into the concept) [practitioner: shhots].

Model. Market: Veo 3.1 for cinematic finish ("24fps native, 4K, broadcast-ready") [practitioner: wavespeed]. Ours: Veo 3.1 with NB2/Pro first frames + character refs; Omni for inserts without faces. Veo timestamp prompting ([00:00-00:02] …) works for multi-beat 8 s takes; we still cut to 1-3 s.

Image prompt.

Cinematic film still, contemporary Tunisia 2026. A Tunisian man in his late 40s, short salt-and-pepper beard, navy work polo, stands on the flat rooftop terrace of his two-storey beige villa in the Sfax suburbs at golden hour, looking at newly installed solar panels. Low sun from camera-right, long shadows, warm bounce from the whitewashed parapet. Shot on ARRI Alexa 35 with a 40mm anamorphic lens, shallow depth of field, oval bokeh on distant olive trees, Kodak Vision3 250D film look, fine grain, natural skin texture with pores, no text.

Motion prompt.

Single continuous shot. Slow push-in from medium to medium close-up over 4 seconds. He exhales and gives a small satisfied nod, eyes staying on the panels; head does not turn. Light breeze moves his polo collar. Handheld-free, smooth dolly, no cuts. No music, no dialogue.

Realism killers. Uncanny faces on head turns (keep faces 3/4, minimal head rotation, <3 s on face) [practitioner: Runway, invideo]; "pleasant cloudy afternoon" flat lighting (name one hard source + direction) [practitioner: OpusClip]; over-saturated synthetic palette (pull saturation 10-15%); letterboxing 2.39:1 inside 9:16 wastes the screen; get the anamorphic feel from bokeh/flare/grade instead [ours].

Post. Film Look chain (§4) with halation 10-20% on highlights, grain 2-3%, gate weave ≤1 px, 24 fps master.

Tunisia. Golden hour over Sfax olive groves, blue hour over Tunis rooftops with the Lac skyline, Sidi Bou Said only if the brand is there (it is a postcard cliché).

3.3 Documentary / "real life" handheld realism

Look. Shoulder-mount or handheld 16 mm/verité: micro-shake, soft focus breathing, occasional missed focus, natural available light, honest desaturated grade, long-ish holds, observational (no eye contact), candid moments, real ambience. Workers doing real tasks.

Why it converts. 2026 trend toward authentic-feeling content; native handheld beats studio production in Reels benchmark sets [practitioner: digitalapplied, Higgsfield doc guide]. For installers it doubles as proof: "this is how we install".

Model. Omni handles "handheld smartphone-style shot" and "natural smartphone zoom" vocabulary well [practitioner: Runware Omni guide]; Veo 3.1 Fast for faces. Our split: Omni for hands/process shots, Veo Fast for faces.

Image prompt.

Candid documentary photograph, 16mm film look. Two Tunisian installers in grey work trousers, safety harnesses and worn gloves bolt an aluminium rail onto a flat concrete roof in Ariana, midday, harsh overhead Mediterranean sun, deep hard shadows, a cylindrical solar water heater and satellite dish in the background, rebar sticking up from an unfinished upper floor next door. Slightly off-centre composition, natural skin texture, sweat on the neck, unposed, no text.

Motion prompt.

Handheld shoulder camera, documentary style, natural micro-shake, no stabilisation. The nearer installer tightens the bolt with a ratchet, two short turns; the camera drifts slightly left and refocuses. Background stays still. Single continuous shot.

Realism killers. Motion "too smooth, no micro-jitter" (ask handheld + add post shake) [practitioner: OpusClip]; hand-tool interaction morphs (tight framing on the tool, gloves hide fingers, cut at 1.5 s) [practitioner: genra]; background extras walking on one leg (no crowds).

Post. Soft Light grain 3-5%, slight chromatic aberration (1 px), saturation -15%, real field recordings.

Tunisia. Grey work trousers and cheap hi-vis, not Western-branded PPE; Isuzu/Peugeot pick-up (blue or white) with ladder; concrete rooftops with rebar and parapets; ambient sound of distant traffic horns, a mosque call far away used sparingly.

3.4 Luxury / premium spot

Look. Restraint: slow moves (dolly-in, low-angle track, slow pan across so highlights travel), single sculpted key, dark marble or matte black sets, specular roll-off, shallow DOF, 15 s default length for luxury [practitioner: Hailuo, perfume prompt corpus]; grade with rich blacks, clean whites, gentle halation; slow-motion mist/fabric.

Why it converts. For premium SMEs (villa solar with storage, smart-home alarms, jewellers, clinics, real estate in Gammarth/La Marsa) it signals price tier; "perceived as AI" hurts premium perception most (-17%), so luxury needs the highest finish [practitioner: digitalapplied].

Model. Veo 3.1 Standard 1080p/4K (cinematic lighting defaults better than Omni) [practitioner: aireiter].

Image prompt.

Luxury product photograph. A slim black hybrid inverter with a single glowing white status LED mounted on a dark travertine wall in a modern Gammarth villa garage, key light a narrow strip softbox from top-left grazing the brushed aluminium edge, anisotropic highlights, deep shadow falloff, polished concrete floor reflecting a soft glow. 90mm lens, f/4, medium-format photograph, restrained, no text.

Motion prompt.

Slow cinematic dolly-in and slight low-angle rise over 5 seconds. The highlight travels smoothly along the brushed edge; the LED breathes once. Everything else perfectly still, no flicker, no deformation.

Realism killers. Blown highlights hiding the product ("large diffused light panel" instead of spotlight); metal-as-plastic ("high dynamic range", "reflecting a studio environment"); flicker (one light, i2v anchor) [practitioner: Hailuo]. Stacking effects (particles + fog + flares) reads cheap.

Post. 4K Veo final or Starlight upscale; halation 10%, grain 1-2%, no chromatic aberration.

Tunisia. Modern villas in Gammarth, La Soukra, Hammamet; Tunisian marble/travertine and olive-wood; avoid Gulf-style gold opulence.

3.5 Drone and establishing aerials

Look. Top-down, orbit, reveal-over-parapet, slow flyover, FPV cinewhoop dive; atmospheric depth haze, light scatter, tiny judder on fast moves, dust/spray near surfaces, wind in vegetation; one continuous move with clear waypoints ("through, down, up") [practitioner: aiolaola, hitpaw].

Why it converts. Instant establishing ("this is your neighbourhood") and the rooftop-array money shot; drone footage of real installs is standard in solar creative.

Model. Veo 3 / Sora give the best photoreal terrain and haze; Kling/Seedance best for fast FPV [practitioner]. Ours: Veo 3.1 Fast i2v from an NB2 aerial first frame; Omni fine for slow top-downs.

Image prompt.

Aerial drone photograph, straight top-down at 60 metres, a dense residential block in Sfax: flat concrete rooftops with parapets, cylindrical solar water heaters, satellite dishes, laundry lines, one roof fully covered by neatly aligned grey-black solar panels facing south. Late afternoon sun from the west, long hard shadows, light dust haze. DJI Mavic 3 photo, realistic, no text.

Motion prompt.

Drone slowly rises and rotates 20 degrees clockwise over 5 seconds, keeping the solar roof centred. Slight wind judder, natural haze, no cuts, roofs and panels keep their exact geometry.

Realism killers. "Game-engine flythrough, rubber-sheet landscape", perfectly clean terrain, oversaturation, cuts inside an FPV move [practitioner: aiolaola]; repeating texture tiling on large roofs/fields [practitioner: OpusClip]; geography nonsense (a medina skyline next to sea cliffs). Famous landmarks drift (avoid close landmark aerials) [practitioner: Sunra].

Post. Keep haze, slight contrast lift, don't sharpen; add rotor-less wind bed. Tunisian drone law restricts real drone use; AI aerials are a real advantage but must not depict identifiable real private properties as "our client" unless they are [ours].

Tunisia. Real rooftops are dense, flat, mixed finished/unfinished, full of water heaters and dishes; not white Santorini cubes. Djerba: menzel domes and palms; Sousse/Monastir: coastline; Tunis: Lac, El Menzah blocks.

3.6 Slow motion and high-speed action

Look. Phantom-style 1000 fps: water droplets, dust puffs, sparks, fabric ripples, strained muscles; real-time setup → slow payoff → release; every micro-detail readable (breath vapour, sweat drop).

Why it converts. Pattern interrupt in the first 2 s; makes an ordinary moment (rain hitting a panel, a door slamming as an alarm fires, a light switching on after a blackout) cinematic.

Model. Veo 3 gives the cleanest high-detail slow-mo; Kling/Seedance for action physics [practitioner]. Ours: Veo 3.1 Fast/Standard with "shot on a Phantom high-speed camera at ~1000 fps"; Omni for simple object slow-mo. Extra slow-down in post: Topaz Apollo at 2x/4x/8x only [official Topaz doc]; never interpolate the whole ad [07 §2.6].

Image prompt.

High-speed photograph, a heavy rain drop hitting the glass of a solar panel on a Tunisian rooftop, crown-shaped splash frozen mid-air, backlit by low sun from behind, dark stormy sky, macro 100mm, f/8, crisp droplets, no text.

Motion prompt.

Extreme slow motion as if shot on a Phantom high-speed camera at 1000 fps. The drop hits, the crown splash rises and collapses, smaller droplets scatter and roll down the glass. Locked-off camera. No stutter, no morphing.

Realism killers. "Slow-mo, cinematic" alone → stuttery smooth-but-fake motion; name the camera and frame rate; use negatives "motion interpolation artifacts, stutter, plastic CG" in Veo's negative prompt; avoid triumphant hero poses [practitioner: aiolaola].

Post. Speed ramps (real-time → slow → real-time) done in Remotion/ffmpeg setpts on 24 fps material look fake below 50%; generate the slow-mo natively, then Apollo 2x if needed [ours + Topaz].

Tunisia. Autumn storms on Tunis rooftops, summer sirocco dust blowing across a Sfax roof, the lights of a villa switching back on after a "coupure". Keep food shots out unless the client sells food.

3.7 Photoreal lifestyle (families, homes, shops)

Look. Natural window light, 35 mm at eye level, medium shots, lived-in interiors with one imperfect detail, candid expressions, practical lamps at night, warm but not orange grade.

Why it converts. Aspiration + identification ("people like me in homes like mine"). Most risky style for AI tells (faces, hands, multi-person).

Model. Veo 3.1 (faces with "particular care, reducing uncanny valley") [practitioner: MindStudio]. Ours: NB2/Pro first frame with character sheet → Veo 3.1 Fast; Omni only for no-face beats (hands on a switch, back of heads, over-shoulder).

Image prompt.

Candid lifestyle photograph, contemporary Tunisia 2026. A Tunisian couple in their 30s in the living room of their apartment in El Menzah on a hot August evening: she has dark curly hair tied back and a light cotton dress, he wears a grey T-shirt; air conditioner running on the wall, sofa with a crocheted throw, ceramic tiles on the floor, tea glasses on a low table. Warm practical lamp light from camera-left, blue dusk through the window. Shot on Sony A7 IV, 35mm f/2, natural skin texture, slight under-eye shadows, unposed, no text.

Motion prompt.

Static tripod shot. She relaxes back into the sofa and lets out a breath; he glances at the air conditioner and smiles slightly. Small natural movements only, no head turns, no hand gestures. Single continuous shot. No dialogue.

Realism killers. Children (not allowed: MENA allow_adult, TikTok minors rule) → imply family with toys, school bag, a drawing on the fridge; multi-person coordinated action (handshakes, passing objects) → one action per shot [practitioner: Runway]; hands → hide or keep simple; mirror/reflection breakdowns (avoid mirrors and glass tables) [practitioner: OpusClip]; stock-photo polish (perfect teeth, catalogue interiors) → "lived-in", "one imperfect detail".

Post. Warm shadows / neutral highlights, grain 2%, softening 0.5 px on skin/hair edges.

Tunisia. See §6. Tunisian interiors: ceramic floor tiles, heavy curtains, salon sofas, framed Quran verse or family photos (don't render text; blur), AC split units everywhere in summer, a "coupure" (power cut) scene lit by phone torches is a strong pain-point opener for solar + battery.

3.8 Before/after and transformation realism

Look. Identical framing, lens and light in both states; only the subject changes (bare roof → PV array; dark villa → lit villa; unguarded shop → camera + keypad); reveal by wipe, match cut, time-lapse or camera move.

Why it converts. Before/after install videos ~2x leads vs static for solar; bill before/after is the #1 hook (use the real bill, composited) [practitioner: webtonic, leadsync].

Model. Market: first+last-frame interpolation (Veo 3.1, Kling). Ours: - Default (house-rule compliant): NB2 generates "before", then an NB2 semantic edit of the same image makes "after" (same geometry guaranteed) → Remotion wipe/slider/match cut. Cheapest and most controllable. - Motion transformation: Veo 3.1 first+last frame (8 s) with both NB2 stills ("the same roof… the same camera position"). Omni supports last frame too, but our house rule is "no last frame on Omni"; propose a rule change only for this style after a test.

Image prompt (before) + edit prompt (after).

Before: Photograph from a neighbour's rooftop, 24mm, of a bare flat concrete roof of a two-storey white-rendered house in Sousse, parapet, solar water heater, midday sun, hard shadows, no text. After (NB2 edit on that image): Keep everything identical (camera, light, shadows, water heater, parapet). Add 12 grey-black monocrystalline panels on aluminium tilt frames facing south in two neat rows, with cable trays to the stairwell. Photoreal.

Motion prompt (Veo first+last).

The same roof, same fixed camera position. Time-lapse over one working day: installers' shadows flicker across as rails and then panels appear row by row; sun moves slightly. Locked-off camera, no cuts.

Realism killers. Geometry drift between states (use edits, not two generations); "morph" transitions that melt the house (cut or wipe instead); fake bills (never AI-generate a STEG bill: scan the client's real one).

Post. Match exposure and grade both halves exactly; add one "real" signal (a phone photo of the actual install).

Tunisia. STEG meter niche on the facade, the STEG bill layout (composited), Prosol Elec context (30% rebate, ~90,000 homes equipped) as copy, not image [official: pv-magazine/ANME].

3.9 Realistic screens and phones in hands

Look. Phone held in a real hand in a real place, screen showing the app/camera feed/bill; slight hand tremor, reflections on the glass, correct perspective.

Why it converts. Alarm/camera installers sell "see your home from your phone"; solar sells "monitor your production in the app". Screen content must be real and legible.

Model / method. AI never renders the UI (gibberish text, drift). NB2 still with a blank, evenly lit (or green) screen → Omni/Veo i2v with minimal motion → Remotion composites the real UI with a 4-corner pin. Keep the phone nearly static relative to the frame (tripod, phone resting in hands on a table) so the corner pin can be keyframed with a few points; or animate the whole plate in Remotion (slow zoom) instead of in the model [ours; screen-replacement practice from stock/mockup tools].

Image prompt.

Over-the-shoulder photograph, a Tunisian woman in her 40s sitting in a car in a Sousse street at dusk holds a smartphone in her right hand, screen plain uniform green, fingers not covering the screen, dashboard and blurred streetlights ahead, iPhone 15 Pro main camera look, natural skin texture, no text, no logos.

Motion prompt.

Static shot. Her thumb stays still at the edge of the phone; she breathes, slight natural hand tremor, streetlights flicker softly in the background bokeh. The phone does not rotate or move out of frame.

Realism killers. Six fingers or fingers through the phone (frame one hand, phone edge-on); moving screen that can't be tracked; green spill on skin (use a dark-grey screen and screen-blend the UI instead of keying) [ours].

Post. Add screen glare/reflection layer in Remotion at 5-10% and a subtle moiré-free brightness; match the UI white point to the scene grade.

Tunisia. Arabic/French UI mixed; the UI is code, so RTL is controlled by us.

3.10 "Shot on iPhone" native-looking realism (UGC-native)

Look. Vertical, front or main camera, 24-26 mm equivalent, deep focus, Smart HDR look (lifted shadows, slightly crunchy highlights), auto-exposure shifts, handheld micro-shake, off-centre framing, real rooms, mixed light (window + LED ceiling), phone-mic room tone, 30 fps.

Why it converts. The dominant winning format: native creator-style Reels +25-40% engagement vs polished; Meta 2024 data cited at 4.2x engagement / 2.8x conversion for UGC-style vs brand creative (vendor-reported); for solar, owner-on-camera and phone testimonials cut CPL 25-40% [practitioner]. AI UGC "realism comes from prompting, not the model": say "shot on iPhone 15", "not a professional photo", "natural imperfections"; avoid studio light, perfect composition, beauty-filter look [practitioner: markbuzz, AdLibrary, atlascloud].

Model. Market: Seedance 2.0 dominates AI UGC; Veo 3.1 handles talking-head UGC with native dialogue. Ours: Omni 720p is fine for this style (resolution and slight softness read as phone), Veo Fast for faces talking. Lip sync to our Gemini TTS Derja voice is a separate problem: prefer voice-over over B-roll, or talking shots <3 s, because Omni lip sync drifts after 6-7 s and nothing on Vertex lip-syncs to external Derja audio reliably [ours + practitioner].

Image prompt.

Ultra realistic iPhone 15 front-camera selfie frame, vertical. A Tunisian man in his early 30s, short fade haircut, trimmed beard, black zip hoodie, on the flat roof of his family house in Ben Arous at 5 pm, solar panels behind him, slightly off-centre, arm's-length angle from a little below, auto-exposed sky slightly blown, real skin texture with visible pores and a little shine, candid mid-sentence expression, not a professional photo, no text.

Motion prompt.

Handheld selfie, chest-up, natural micro-shakes from holding the phone at arm's length. He talks animatedly, small head movements, glances once at the panels behind him and back to the lens. Auto-exposure adjusts slightly as the phone tilts. Single continuous shot.

Realism killers. Over-politeness and "ad voice" in the script (write for Derja speech with fillers, micro hesitations) [practitioner: atlascloud]; beauty-filter skin; perfect lighting; captions generated by the model (house rule: text in Remotion); the fake-testimonial trap (§5).

Post. No film grain; instead light sensor noise (luma noise 3-4), slight oversharpening is authentic to phones, chromatic aberration 1 px, subtle shake from ffmpeg crop expressions, 30 fps master, phone-mic EQ on voice.

Tunisia. The strongest authenticity move: have the real installer film 5-10 s of himself on his phone and use AI only for B-roll around him. AI "owner" should be a spokesperson character, not an impersonation.

3.11 Other photoreal sub-styles that perform

a) CCTV / security-cam POV (alarms & cameras). Wide fisheye 2.8 mm look, high angle, IR night mode (monochrome, blown IR highlights), timestamp overlay (Remotion), low frame rate feel. The product is the camera, so the footage is the demo. Prompt: "Security camera footage from a 2.8mm wide-angle CCTV camera mounted under the eave of a Tunisian villa at night, infrared black-and-white, slight fisheye distortion, a gate and parked Peugeot 208 in the courtyard". Motion: "Static camera. A cat walks along the wall; motion is slightly choppy like 15 fps surveillance video". Post: drop to 12-15 fps (fps=15), IR bloom, compression blockiness, scan-line/lens distortion (Higgsfield added CCTV authenticity in Resolve this way) [practitioner]. Never stage a crime by a realistic person breaking in presented as real footage (fabricated event): use a dramatization label or show the alarm/deterrence, not the burglary [ours].

b) Tradesperson POV / process ASMR. Chest-mount GoPro POV of gloved hands crimping MC4 connectors, drilling, clicking a panel clamp; tactile sounds. Hands are the risk: gloves, tools filling the frame, 1-1.5 s cuts. Omni is good here (object physics) [practitioner: invideo Omni review].

c) Time-lapse / hyperlapse. Sun arc over a rooftop array, shadow sweep, day-to-night villa lighting up from its own battery. Prompt "locked-off time-lapse over one day", or build it from NB2 edits (same image, 4 times of day) + Remotion crossfades: fully controllable and cheap [ours].

d) Vox-pop / street-interview look. Handheld interviewer mic in frame, passers-by in Avenue Habib Bourguiba. Performs as native content but is the highest compliance risk (fake real people giving opinions). Use only with real people, or clearly scripted "characters" with disclosure; never fake opinions about savings [ours + policy].

e) Pain-point realism. Sweating in a heatwave with the AC off, a "coupure" in a dark apartment lit by phones, a shop owner fanning himself. Problem shots don't need the product and are safer for AI (no brand fidelity needed).

4. Finishing to 10/10 (photoreal-specific layer on top of 07 §2.5)

4.1 Per-style finishing matrix [practitioner values + ours]

Rules: pick one master fps per ad; grain goes after upscale and before encoding; platforms re-encode, so grain that is visible on the desktop disappears or turns to blocks on Reels, keep it subtle and export at a high bitrate [ours]. Sound is the biggest realism lever (07 §3). AI clips' native audio is scratch only.

4.2 ffmpeg building blocks (test values, tune per ad) [ours, filters per ffmpeg docs / zayne.io]

# Colour: LUT at partial strength + saturation trim
lut3d=file=looks/vision3_250d.cube, eq=saturation=0.88:contrast=1.03
# Split-tone: warm shadows, cooler highlights
colorbalance=rs=0.03:bs=-0.03:rh=-0.01:bh=0.02
# Halation (film): highlight pass, blurred, tinted red-orange, screened back
split[a][b];[b]curves=all='0/0 0.75/0 1/1',gblur=sigma=14,colorchannelmixer=rr=1:gg=0.35:bb=0.1[h];[a][h]blend=all_mode=screen:all_opacity=0.15
# Anamorphic streak (luxury/night): horizontal-only blur on highlights, tinted blue
[b]curves=all='0/0 0.85/0 1/1',gblur=sigma=40:sigmaV=0.1,colorchannelmixer=rr=0.3:gg=0.5:bb=1
# Grain: luma only, temporal (≈ 6-10 at 1080p cinematic; lower at 720p)
noise=c0s=8:c0f=t+u
# De-crisp AI edges
gblur=sigma=0.5            # or unsharp=5:5:-0.3
# Chromatic aberration (UGC/doc only)
rgbashift=rh=1:bh=-1
# Vignette
vignette=PI/5
# Gate weave / handheld micro-shake (crop with moving offset, then scale back)
crop=iw-12:ih-12:6+2*sin(n/7):6+2*cos(n/11),scale=1080:1920
# Surveillance cadence
fps=15
# Deliverable (house rule)
format=yuv420p, -colorspace bt709 -color_primaries bt709 -color_trc bt709

Note: lenscorrection can add slight barrel distortion for wide/GoPro/CCTV looks (k1/k2 around ±0.02; check the sign visually). Real LUTs (Vision3 emulations) beat hand curves for the cinematic grade.

4.3 Upscale and interpolation

• Veo: native 1080p/4K only on 8 s generations. Omni: 1080p/4K are upscales (billing tests), so Topaz Starlight Precise 2.5 on the kept 1-3 s is usually better value than paying Omni 4K [practitioner].
• Starlight Precise 2.5 for realism (faces, fabrics, logos) vs Astra 2 (creative, invents detail: never on faces or products) [practitioner: invideo Topaz].
• Apollo for 2x/4x/8x slow-mo; Chronos for non-integer retimes; never interpolate a whole ad (rubbery, ghosting) [official Topaz + 07].
• Export intermediates in ProRes 422 before grading to avoid stacking compression [practitioner: Chase Jarvis].

5. Platform rules for realistic AI people (Meta, TikTok, law)

Meta (Facebook/Instagram). - "AI info" label is applied automatically for Meta's own gen-AI tools and, since the June 2026 update, for third-party tools via industry signals such as C2PA; advertisers can self-disclose at ad level. For photorealistic AI people the label can move next to "Sponsored" in the feed [practitioner: coinis, auditsocials]. - Undisclosed AI content was reported at ~14% of ad rejections in 2026; escalation: ad killed + strike → 24 h hold → suspension risk [practitioner: coinis, vendor-reported]. - Spokesperson vs fake customer: an AI avatar delivering a scripted brand message is fine; "Sarah, a verified buyer" is deceptive advertising regardless of disclosure [practitioner: cinerads]. Unsupported performance claims (e.g. "zero bill") are rejected regardless of creator type. - Official pages: Meta Business Help Center AI disclosure, Transparency Center "Labeling AI content".

TikTok. - AIGC label mandatory for realistic-appearing people or scenes; ads use the "This ad contains AI-generated content" toggle (cannot be changed after submission; duplicating a campaign resets it); C2PA auto-labelling; escalation from label → removal → account strikes [practitioner: cinerads, auditsocials, storrito]. - Removed outright even if labelled: synthetic likenesses of anyone under 18, realistic depictions of non-public adults without consent, fabricated authoritative sources/crisis events, false public-figure endorsements.

Google side. Every Veo/Omni output carries an invisible SynthID watermark [official]. Whether Vertex video outputs also carry C2PA metadata was not confirmed in this research; assume platforms can detect AI and self-disclose.

Law. FTC fake-review/testimonial rule (Oct 2024) and AI-disclosure guidance; New York "synthetic performer" disclosure law (June 2026); EU AI Act Art. 50 labelling from 2 Aug 2026 (applies if we ever target EU diaspora) [practitioner: cinerads]. No Tunisian AI-ad law found; platform rules are the binding constraint.

Does the label hurt? MediaScience/MediaPet/Ehrenberg-Bass (June 2026, n=900): unaided recall 54-61% and recognition 87-92% across labelled/unlabelled groups, i.e. negligible impact; "consumers are accepting of AI if they know… provided the ad is good" [practitioner: Marketing Dive]. What hurts is perceived AI-ness and low craft (NIQ).

Our compliance checklist [ours]. 1. Tick the AI disclosure on every photoreal ad (Meta + TikTok). 2. No realistic minors. Families = adults (couples, grandparents, adult siblings); kids implied by objects. 3. No AI person named or captioned as a real customer; testimonials only from real, consenting customers filmed on phones. AI people = "characters" or the brand spokesperson. Add "تمثيل / représentation" small on-screen where a scene dramatizes a situation (our choice; not a platform requirement). 4. No likeness of real public figures or identifiable real private people; no real private houses shown as clients' unless they are. 5. No fabricated events that could read as news (burglary "caught on camera", STEG announcements). 6. Savings numbers must be the installer's substantiated numbers.

6. Tunisian authenticity kit (all styles)

Why it matters. T2I models show "Visual Orientalism": Eastern countries rendered through traditional/exotic symbols (4.02 cultural vs 0.20 political/modern symbols per image) and non-Anglo countries exoticised; adding "contemporary"/"modern" temporal markers partially counters it [arXiv 2511.22931]. Out of the box, "Tunisian house" gives Sidi Bou Said blue doors, "Tunisian man" gives a chechia and jebba.

Prompt anchors that work (use concrete nouns, not "Tunisian style") [ours, grounded in sources where noted]. - Always: "contemporary Tunisia 2026", city + neighbourhood (Ennasr, El Menzah, Ariana, Ben Arous, La Marsa, Sfax El Ain road, Sousse Khezama, Djerba Houmt Souk). - Houses: two-storey rendered concrete villa, beige/white/pale-yellow render, fer forgé window grilles, aluminium shutters, flat roof terrace with parapet, cylindrical solar water heater (PROSOL has equipped ~900,000 m² since 2005), satellite dishes, rebar left for a future floor, laundry lines. Djerba: whitewashed menzel with domes, palms. Medina: narrow lanes, studded doors, only when the client is in the medina. - Solar specifics: grey-black mono panels on aluminium tilt frames on flat roofs facing south; wall-mounted inverter; STEG meter niche on the facade. Prosol Elec: 30% rebate, free STEG inverter, credit up to TND 3,000/kW, ~300 MW in ~90,000 homes [official via pv-magazine/ANME]. - Streets and vehicles: white louages with a red stripe (inter-governorate), blue (within a governorate), yellow (rural) [Wikipedia]; yellow Tunis taxis; Peugeot/Renault/Kia compacts, Isuzu pick-ups; TGM and green light metro in Tunis. - People: Mediterranean North African range of skin tones (fair-olive to brown), dark hair; men mostly jeans/chinos, polos, hoodies, short fades and trimmed beards; women both uncovered hair and hijab, modern clothing; traditional dress (jebba, chechia, sefsari) mainly for older people, weddings and religious occasions [practitioner: tunisiaonline, Wikipedia Safseri]. Workers: grey work trousers, cheap hi-vis, caps. - Light: hard high Mediterranean sun, deep shadows, bright bounce off white walls; summer haze; golden hour warm but not orange. - Avoid (clichés and wrong countries): Gulf thobes/ghutra, Moroccan djellabas with pointed hood, babouches, zellige everywhere, tagines, lanterns, camels/desert (unless the story is in the south), belly dance, the Hollywood "yellow/sepia filter" on MENA scenes, Sidi Bou Said blue-and-white as a generic "Tunisia". - Text in scene: Arabic/French signage, licence plates, shop fronts, bills → gibberish. Keep them out of focus or out of frame and composite real text in Remotion (house rule). - Best lever: real reference photos (client's roof, street, van, uniform, the owner's face with consent) fed to NB2 as references. Up to 14 refs per image.

7. Model routing cheat sheet (Vertex only) [ours]

8. Prompt skeletons

NB2 first frame (photoreal). [Shot type + camera body + lens + aperture]. [Subject with age, look, wardrobe, one imperfection]. [Action frozen mid-moment]. [Specific Tunisian place + time + season]. [Light: source, direction, quality, bounce]. [Film stock or phone look, grain]. [Realism cues: natural skin texture, pores, unposed, lived-in detail]. No text, no logos. Words to avoid: perfect skin, flawless, airbrushed, beautiful, masterpiece, 8K, Octane, trending [practitioner: Luno]. One film stock per prompt.

Motion (Omni: motion only; Veo: add "no music, no dialogue"). [Continuity: single continuous shot]. [Camera move + speed + duration]. [ONE subject action, small, with weight]. [Secondary micro-motion: breeze, dust, breath]. [Guards: identity/shape stays identical, no morphing, no flicker, no head turn]. Lead with the camera phrase (Omni reads the opening as the primary direction) [practitioner: Runware]. Prefer slow, measured moves; one main action; test 3-5 s first [practitioner: Runway].

9. Recommendations for Lead Up

1. Add a Veo 3.1 route to the engine next to Omni (Vertex veo-3.1-fast-generate-001 for takes, veo-3.1-generate-001 for finals, 9:16, personGeneration=allow_adult, seed stored per take).
2. Build a photoreal style kit per client: cast sheets (2-3 characters), 3-5 location plates from real photos, one LUT, the finishing preset from §4.1, and a disclosure checklist.
3. Default photoreal lead ad = iPhone-native + documentary + real bill/real install clip, with 1-2 cinematic shots (drone reveal, macro panel) as production value. Cinematic-only films for brand/retargeting.
4. Pilot test (after script approval): one 30 s solar ad in route B (§2.3), measure keep rate per model and time, then decide whether Omni faces are acceptable at 1.5 s cuts.
5. Open questions to verify with a small test: Nano Banana Pro availability on our Vertex project; C2PA presence in Vertex video outputs; whether Veo i2v with a first frame honours asset references on Vertex; Omni last-frame use for before/after (needs owner sign-off to relax the house rule).

Sources

Official / primary - https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-veo-3-1 - https://ai.google.dev/gemini-api/docs/veo - https://ai.google.dev/gemini-api/docs/omni - https://ai.google.dev/gemini-api/docs/pricing - https://cloud.google.com/blog/products/ai-machine-learning/ultimate-prompting-guide-for-nano-banana - https://blog.google/innovation-and-ai/technology/ai/veo-3-1-ingredients-to-video/ - https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/ - https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/generate-videos-from-an-image - https://docs.cloud.google.com/vertex-ai/generative-ai/docs/video/video-gen-prompt-guide - https://docs.topazlabs.com/video-ai/how-to-guide/using-apollo-to-create-a-smooth-slo-mo-output - https://ffmpeg.org/ffmpeg-filters.html - https://www.facebook.com/business/help/1010479435004531 - https://transparency.meta.com/governance/tracking-impact/labeling-ai-content/ - https://www.tiktok.com/creator-academy/en/article/ai-generated-content-label

Model comparisons, pricing, reviews - https://aireiter.com/blog/gemini-omni-flash-vs-veo-3-1 - https://wavespeed.ai/blog/posts/omni-flash-vs-veo-sora-seedance/ - https://www.mindstudio.ai/blog/seedance-2-vs-veo-3-1-comparison - https://magicshot.ai/blog/kling-vs-veo-vs-seedance-comparison - https://invideo.io/blog/gemini-omni-flash-review/ - https://www.buildfastwithai.com/blogs/gemini-omni-flash-review-google-ai-video-model-2026 - https://www.digitalapplied.com/blog/gemini-omni-multimodal-video-creation-guide - https://www.atlascloud.ai/blog/tips/gemini-omni-flash-1.1-pricing - https://www.eesel.ai/blog/gemini-omni-1-1-flash-pricing - https://costgoat.com/pricing/google-veo - https://www.cined.com/google-veo-3-1-ingredients-to-video-update-adds-native-vertical-format-4k-upscaling-and-enhanced-character-consistency/ - https://www.banani.co/blog/nano-banana-pro-vs-nano-banana-2 - https://vercel.com/ai-gateway/models/veo-3.1-fast-generate-001 - https://docs.litellm.ai/docs/providers/vertex_ai/videos

Prompting and craft (practitioner) - https://runware.ai/docs/models/google-gemini-omni-flash/guides/cinematic-prompting - https://www.atlabs.ai/blog/ultimate-gemini-omni-flash-prompting-guide-2026 - https://runway.com/resources/tips-make-realistic-ai-videos - https://lumalabs.ai/news/cinematic-ai-video-prompts - https://www.lunostudio.ai/academy/nano-banana-photorealism - https://hailuoai.video/pages/knowledge/specular-highlights-luxury-ai-video-guide - https://blog.picassoia.com/veo-3-1-for-product-launch-videos - https://github.com/LichAmnesia/awesome-ad-video-prompts - https://prompts.aiolaola.com/p/drone-fpv - https://prompts.aiolaola.com/p/sports-slowmo - https://edimakor.hitpaw.com/video-editing-tips/drone-shot-ai-prompt.html - https://morphic.com/ai-glossary/handheld-shot - https://higgsfield.ai/blog/ai-documentary-videos-2026 - https://higgsfield.ai/blog/ai-commercial-youtube-guide - https://www.atlascloud.ai/blog/tips/how-to-make-ai-ugc-ads - https://markbuzz.in/how-to-create-hyper-realistic-ai-ugc-images-that-actually-look-real-complete-prompt-guide-for-2026/ - https://adlibrary.com/guides/ai-prompting-guide-ugc-content-creators - https://invideo.io/faq/why-do-ai-generated-videos-look-plasticky-and-fake-and/ - https://invideo.io/faq/does-adding-film-grain-and-blur-to-ai-generated-video/ - https://invideo.io/faq/how-do-you-use-a-character-reference-sheet-to-keep-ai/ - https://invideo.io/blog/topaz-video-enhancement-models/ - https://apipass.dev/blogs/master-character-consistency-nano-banana-2-api - https://sunra.ai/blog/make-ai-video-look-less-ai - https://www.opus.pro/blog/ai-slop-aesthetic-12-tells - https://genra.ai/blog/why-ai-videos-look-fake-how-to-fix - https://zayne.io/articles/vintage-camera-filters-with-ffmpeg - https://chasejarvis.com/blog/veo-3-faqs/ - https://www.vidhex.ai/blog/make-ai-skin-look-real/ - https://dreamina.capcut.com/ai-video/how-to-use-start-and-end-frame-generators

Case studies - https://threadreaderapp.com/thread/1932893260399456513.html (PJ Accetturo, Kalshi) - https://www.marktechpost.com/2025/06/14/ai-generated-ad-created-with-googles-veo3-airs-during-nba-finals-slashing-production-costs-by-95/ - https://shhots.ai/blog/best-ai-commercials/

Ad performance and platform policy - https://www.digitalapplied.com/blog/instagram-reels-ads-2026-creative-strategy-guide - https://www.digitalapplied.com/blog/ai-ad-creative-benchmark-2026-ctr-roas-data - https://www.webtonic.io/blog/solar-energy-ad-creative-statistics - https://leadsync.me/blog/solar-facebook-ads/ - https://www.marketingdive.com/news/ai-disclosure-labels-dont-hurt-ad-performance-heres-what-the-numbers-say/822711/ - https://nielseniq.com/global/en/news-center/2024/niq-research-uncovers-hidden-consumer-attitudes-toward-ai-generated-ads/ - https://coinis.com/blog/meta-ai-content-labeling-facebook-instagram-ads-2026 - https://www.cinerads.com/blog/ai-ugc-facebook-ad-policy - https://www.cinerads.com/blog/tiktok-ai-content-policy - https://www.cinerads.com/blog/ai-ad-disclosure-requirements - https://www.auditsocials.com/blog/tiktok-ai-content-disclosure-rules-2026

Tunisia and representation - https://www.pv-magazine.com/2024/09/03/tunisia-allocates-121-million-for-residential-solar-rebates/ - https://anme.anme.com.tn/en/content/residential-solar-energy - https://sdgs.un.org/partnerships/prosol-solar-programme - https://en.wikipedia.org/wiki/Louage - https://en.wikipedia.org/wiki/Safseri - https://www.tunisiaonline.com/traditional-tunisian-clothing/ - https://arxiv.org/html/2511.22931 (Visual Orientalism in AI image models)