Photorealistic podcast

Podkast format izveden sa fotorealističnim ljudima. Razlika u odnosu na osnovni modul nije u strukturi nego u produkciji: kad lice mora da izgleda kao stvaran čovek, sve se lomi na konzistentnosti lika, teksturi kože i izgovoru. Ova kombinacija se najviše prodaje klijentima jer izgleda kao snimljeno, a pravi se za sat vremena.

Write concept, script and explanations in the user's language. **Write all generation prompts in English**, because the models perform better and the user pastes them unchanged.

What it earns, honestly

Ovde se pada na tri stvari, ne na strukturi. Lice koje se menja između kadrova, koža koja izgleda kao plastika, i izgovor koji odaje stranca. Nijedna od te tri se ne rešava boljim skriptom. Rešavaju se zaključavanjem lika pre prve generacije videa, eksplicitnim traženjem teksture kože, i pravilnim modelom za govor.

Why it persuades

• Fotorealizam nosi autoritet koji nijedan stilizovan lik ne može. Tvrdnja iz usta stvarnog čoveka prolazi kao informacija.
• Klijenti kupuju ovo bez objašnjavanja. Ne moraš da braniš estetiku, samo isporučiš.
• Jedan set i dva zaključana lika daju trideset klipova, pa je cena po reklami najniža u biblioteci.
• Format prolazi kao sadržaj, pa hladna publika ne beži pre nego što padne prva tvrdnja.

Stage 1 — Interrogate the product

Ask what the user is selling, at what price, to whom, and what the single claim is. Do not proceed on assumptions. If they already gave enough context, answer for them and show it for confirmation rather than asking again.

The governing rule: the concept is derived FROM the product, never applied TO it. Quality test before delivering: would this same concept work with a competitor's product? If yes, rewrite.

Stage 2 — Pick the style variant

Cela vrednost kombinacije je u tome što gledalac ne pomisli na AI. Čim se doda bilo kakva stilizacija, prednost nestaje i bolje ti je da uzmeš osnovni podkast modul i biraš svesno.

If the user does not choose, recommend one and say why in one line.

Stage 3 — Script the beats

Reakcija sagovornika je najvredniji kadar u ovoj kombinaciji. On je jedini dokaz da su dva lika u istom prostoru.

Write the actual lines verbatim in the user's language. Do not describe what is said.

Stage 4 — Five hooks against one body

Always deliver five opening variants. When hold rate is fine and hook rate is low, only the first three seconds get regenerated, which is five ads for the price of one production.

Structures that hold for this format:

• „Niko neće da kaže ovo naglas, ali...“
• Sagovornik prekida: „Čekaj, to nije tačno.“
• „Devedeset posto ljudi u ovoj branši radi ovo pogrešno.“
• Duga pauza pa tvrdnja
• „Prestao sam da radim X i prihod se udvostručio.“

The hook must be a single visible thing. "Curiosity" is not a hook.

Stage 5 — Production

Order of work, do not deviate

1. Skupi ili izgeneriši referentne fotografije lika Za Soul ID treba dvadeset i više fotki: različiti uglovi, različiti izrazi, isto svetlo, bar jedna cela figura. Trening traje nekoliko minuta i radi se jednom.
2. Generiši široki kadar On postavlja set, svetlo i oba lika odjednom.
3. Zaključaj oba lika i set Reference Element ili Soul ID. Bez ovoga se lice menja već u drugom kadru, i to je greška broj jedan u ovoj kombinaciji.
4. Generiši krupne planove iz reference Kopiraj rečenicu o svetlu i visini kamere doslovno.
5. Video, pa provera izgovora Ako srpski zvuči kao stranac, glas se pravi odvojeno pa se lip sync radi preko.

Model routing

Shot	Image	Video
Referentni lik	Nano Banana Pro ili Soul ID trening	nema
Široki kadar	Nano Banana Pro	Kling 2.5, minimalan pokret
Govornik, srpski	Nano Banana Pro	Google Omni Flash
Govornik, engleski	Nano Banana Pro	Veo 3.1
Reakcija bez reči	Nano Banana Pro	Kling 2.5

On-camera Serbian goes through Google Omni Flash. Other models carry a foreign accent local audiences detect within half a second. If pronunciation still fails, generate the voice separately and apply lip sync with an audio-driven tool over the finished image rather than accepting a foreign-sounding take.

Locked parameters

Every prompt carries these, explicitly or by clear implication. Changing one means generating a different format.

Prompt skeletons

REFERENTNI LIK (Nano Banana Pro)

A photorealistic portrait of [PERSON DESCRIPTION: age, build, hair, facial hair, clothing], shot straight on at eye level against a plain neutral background, even soft light with no strong shadows, neutral expression, eyes open and looking into the lens. Natural skin texture with visible pores, faint blemishes and fine lines, no smoothing, no beauty retouching. 50mm lens, sharp focus on the eyes. Passport-like reference image intended for character locking.

**KADAR 1** (široki, definiše set)

Two people seated at a podcast table facing each other slightly off-axis, large dynamic microphones on boom arms between them. Behind them a dark acoustic panel wall with a single [COLOUR] light strip glowing. [SPEAKER A] on the left, [SPEAKER B] on the right. Static camera at seated eye level, 35mm, both subjects in frame, key light from camera left and a practical source behind. Contrastful slightly cool grade with fine digital grain. Natural skin texture with visible pores and faint blemishes, no smoothing. Looks like a frame pulled from an hour-long recording, not a commercial.

**KADAR 2** (govornik krupno)

[SPEAKER] seated at the podcast table, framed from the chest up, slightly off-centre with the microphone entering the lower third of frame. The same dark acoustic panel wall and [COLOUR] light strip behind. Mid-sentence, mouth slightly open, one hand raised. Static camera at seated eye level, 50mm, key light from camera left and a practical source behind, identical to the wide shot. Contrastful slightly cool grade with fine digital grain, natural skin texture with visible pores and faint blemishes, no smoothing.

**VIDEO, govor** (Omni Flash za srpski)

[FIRST FRAME] speaks. In [LANGUAGE], casually and mid-thought with one small natural stumble: "[LINE]". The head moves naturally with the speech, hands occasionally entering the bottom of frame, shoulders shifting slightly. The camera is completely static and does not move. Audio: close-mic voice with light room tone, no music. A single uninterrupted take filmed in one breath.

**VIDEO, reakcija** (Kling 2.5)

[FIRST FRAME] holds. The listener [leans back and exhales half laughing / shakes his head slowly / opens his mouth to interrupt and stops]. He says nothing. The camera is completely static. Audio: light room tone and one faint chair creak, no music. A single uninterrupted take filmed in one breath.

### Two phrases that are not optional

Every video prompt closes with `A single uninterrupted take filmed in one breath`, or the models
insert hidden cuts. Never write `zoom`; write `the camera travels toward`. Never name a phone or
camera, or the model draws one into the shot.

---

## Worked example — Proizvod: konsalting za male firme, 800 evra mesečno. Teza: većina firmi ne gubi novac na troškovima nego na neodlučivanju.

**PRIMER · KADAR** (Nano Banana Pro)

Two people seated at a podcast table facing each other slightly off-axis, large dynamic microphones on boom arms between them. Behind them a dark acoustic panel wall with a single warm amber light strip glowing. A man in his forties with greying short hair in a dark shirt on the left, a woman in her thirties in a rust knit sweater on the right. Static camera at seated eye level, 35mm, both subjects in frame, key light from camera left and a practical source behind. Contrastful slightly cool grade with fine digital grain. Natural skin texture with visible pores and faint blemishes, no smoothing.

**PRIMER · VIDEO** (Omni Flash)

The frame speaks. In Serbian, casually and mid-thought with one small natural stumble: "Firme ne propadaju zato što troše previše. Propadaju zato što... zato što niko ne donese odluku tri meseca." The head moves naturally with the speech, one hand raised, shoulders shifting slightly. The camera is completely static. Audio: close-mic voice with light room tone, no music. A single uninterrupted take filmed in one breath.

Use this to calibrate how much detail belongs in each bracketed slot.

---

## Failures and repairs

**Lice se menja između kadrova**
→ Zaključaj lik pre prve video generacije. Ovo je greška broj jedan u ovoj kombinaciji i ne rešava se opisom.

**Koža izgleda kao plastika**
→ Traži pore, sitne nepravilnosti i fine linije, i nikad ne piši reč cinematic. Podrazumevano lepšanje je najveći AI tell.

**Srpski zvuči kao stranac**
→ Omni Flash. Ako i dalje ne valja, glas se generiše odvojeno pa lip sync ide preko gotove slike.

**Oba lika liče jedno na drugo**
→ Eksplicitno različit uzrast, građa, kosa i odeća. Model sam po sebi teži ka sličnim licima.

**Zvuči kao narator**
→ Traži `casually and mid-thought with one small natural stumble`.

**Mikrofon se deformiše**
→ Statična kamera i zaključan set. Mikrofon je najosetljiviji objekat u kadru.


**Three-attempt rule.** If the same problem survives two generations, change the approach rather
than the prompt: add a reference image, halve the shot, or split the action into two generations
joined in the edit.

## Assembly

- **Redosled** Hook preko širokog, govornik, reakcija, poziv. Do 28 sekundi.
- **Rezovi** Tvrdi, na rečenici. Rez na reakciju posle tvrdnje je najjači.
- **Titlovi** Reč po reč. Prve tri sekunde moraju da rade i bez zvuka.
- **Zvuk** Bez muzike ispod govora. Ambijent sobe je dovoljan.
- **Serija** Isti set i isti likovi, nova teza, novi klip. Trideset klipova iz jedne postavke.

---

## Deliver

KONCEPT:      [one paragraph, what happens and why it comes from this product]
VARIJANTA:    [which style, and why]
SCENARIO:     [beat by beat, with verbatim lines]
PET HOOKOVA:  [five openings against the same body]

Then every prompt in English, complete and pasteable, labelled by shot. Then one line on what to
check first in the result.

## Never

- Deliver prompts without a concept behind them.
- Deliver a concept that would survive a product swap.
- Return prompt fragments or advice about what the user should edit. Full prompts only.
- Recommend a style variant this format does not support.
- Let the word "zoom" or a named recording device into a prompt.
- Substitute a model when Serbian speech is on camera.