# Production grammar **Production Grammar Teardown — a 26.5s streetwear promo** Reverse-engineered as a film-director + AI-video-technical-director document. Purpose: reproduce the *grammar* (shot logic, camera behavior, pacing, assembly) on your own character, environment and script — never the source character, wardrobe, IP or dialogue. **0. Verified technical specs (ffprobe)** |Property |Value | |-----------|------------------------------------------------------------------------------| |Duration |26.53s | |Resolution |774×576 (~4:3) | |Codec / fps|H.264, 30fps | |Structure |7 shots = 6 animated 3D shots + 1 still product photo | |Audio |Full voiceover throughout + lo-fi synthwave bed (~95–100 BPM), one “scoff” SFX| The 4:3 frame is a *design decision*: it pushes the “early-2000s game cutscene” association. Keep it (aspect 4:3 in generation) or fake it (letterbox a 16:9 output). Do not default to 9:16 — that would kill the retro read. **1. Shot structure (timestamps, framing, camera, blocking, acting)** |#|Time |Frame |Camera |Blocking / acting |BG activity | |-|----------------|----------------------------------------------------------------------------|---------------------------------------------|-----------------------------------------------------------------------------------------------------------------------------------|---------------------------------------------------| |1|0:00–0:02 (2s) |MCU, eye-level, slightly low, subject **right of center** leaning on railing|Slow **push-in** |Forearms on rail, direct eye contact; **open palm → fist → fist pulled to chest**, each gesture landed on a word |Cloud-sea, blocky towers | |2|0:02–0:11 (9s) |**Wide / full body**, low-ish angle, camera back on balcony |Static + near-imperceptible **handheld sway**|Weight shifts, shrugs, open-hand “explaining” gestures, head shakes; hands never read as random — each gesture attaches to a clause|Cityscape, tower with bridge strut, clouds drifting| |3|0:11–0:18 (7s) |MCU, tighter, 3/4 side angle |Slow lateral **drift right** |**Prop intro** (ceramic gourd in right hand), smug smirk, head tilt, brow raise on punchline |Same cloud-sea | |4|0:18–0:22 (4s) |**MS, waist-up**, facing camera |Slow push-in |Fist beat on chest → double thumbs → points at shirt graphic; the shirt becomes readable for the first time |Slightly defocused towers (shallow DOF) | |5|0:22–0:23 (1s) |**CU face**, static |Locked-off |Intense deadpan stare into lens, chin slightly down |Blurred railing | |6|0:23–0:26 (3.5s)|Product still (photo) |None |Real garment on model, neutral pose |Plain white studio | **2. Cinematography + editing grammar (transferable)** **Shot-size ladder as narrative.** The piece is one escalating argument, and the lenses follow it: MCU (hook) → WS (context: “let me show you the situation”) → MCU (intimate anecdote) → MS (payoff: point at the thing) → CU (CTA) → product still (transaction). Your reconstruction should keep the *ladder*, not the exact sizes — the ladder is what makes a 26s ad feel like a short film. **Camera behavior is a vocabulary, not a set of SFX.** Almost nothing is static: - *Slow push-in* = “leaning in to tell you something” (shots 1, 4). - *Handheld sway* = “casual, off-the-cuff” (shot 2). - *Lateral drift* = “time passing, storytelling” (shot 3). - *Locked-off CU* = “stop, this is the point” (shot 5). **Composition lock.** Subject consistently on the **right third**, eye contact straight into the lens. The gaze-to-camera is the single strongest stylistic device — it converts a monologue into a confrontation with the viewer. Background keeps the subject separated: sea of clouds, distant towers, strong rim light. **Editing rhythm.** - 6 cuts in 26.5s; average shot ≈4.4s; longest 9s (context), shortest 1s (CTA slam). - Every cut lands **on a punch word or a caption change** — the caption is the metronome. - Cut 1 (MCU→WS, 0:02) is a **jump cut** in scale used to “widen the argument”. - Shots 4→5→6 is a compression: MS → CU → product, accelerating to the CTA, then a **hard cut into** **a different world** (3D → real photo) that says “the thing in the story is the thing you buy”. **Dialogue pacing.** Deadpan, fast, meme-cadence: 2–6 word phrases, short holds, occasional dead-air beat for a joke (a one-word answer … *scoff* … the punchline). The VO never stops; the music never stops; there are no empty seconds. Captions chunk *with* the phrases, not with the sentences. **3. Visual design vs. cinematography/editing** **Belongs to visual design (the “look”)** — reproduce these via style anchors / art direction: - Stylized PS2/PS3-era game-CGI render: low-poly blocky architecture, slightly stiff facial rig, sharp hair silhouette, soft resolution. - Character design: a distinctive silhouette in high-contrast streetwear, with a graphic tee as product placement. - Environment: high-altitude balcony above a **sea of clouds**, minimalist towers, haze depth. - Lighting: **golden-hour sun from screen-left/back-left**, strong warm rim light on hair/shoulders, long shadows on the floor, high contrast with cold blue sky. - Palette: red / gold / white vs. blue-grey; warm grade. - Typography: white bold sans, centered low third, soft black drop shadow. - 4:3 frame. - **IP note:** design a character of your own. Never prompt a licensed character or a recognisable look from an existing franchise. **Belongs to cinematography/editing (the “grammar”)** — reproduce these via shot design + post: - Shot-size ladder MCU→WS→MCU→MS→CU→product. - Camera behavior vocabulary (push-in = intimacy, sway = casual, drift = storytelling, locked = point). - Thirds composition + direct address to lens. - Cut-on-punch-word rhythm; jump cut in scale; hard cut 3D→real at the CTA. - Caption timing synced to phrases; caption-as-metronome. - VO pacing with deliberate pause beats; music constantly under VO; single accent SFX. **4. Higgsfield tool/model map** |Component |Tool |Model |Why | |------------------------------------|-----------------------------------------------------------------------------------------------------------------------|---------------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------------------------------------------------------| |Original character anchor |`higgsfield_generate_image` |`soul_cast` (primary) or `nano_banana_2` (stylized cartoon/3D look) |Reusable identity for later video shots; avoid gpt_image/photo-real human models for the game-CGI look | |Character consistency across shots |persistent element registration (`higgsfield_element`) |— |Same character in all clips without identity drift | |Environment keyframe |`higgsfield_generate_image` |`soul_cinematic` / `soul_cinema_1_5` |Cinema-grade concept/still for the balcony + cloud-sea | |Talking shots (1, 3, 4, 5) |`higgsfield_generate_video` |`seedance_2_5` (`params.mode: "omni_reference"`, 4:3) |Multi-shot continuity, up to 30 refs, native audio, deepest reference capacity; strongest default | |Motion/animation-heavy wide shot (2)|`higgsfield_generate_video` |`minimax_h3` |Best motion/animation model in the stack; 2K; start_image/end_image; audio refs for lip-sync; 4:3 supported| |Product still (6) |`higgsfield_generate_image` |`text2image_soul_v2` (realistic model wearing the merch) or `recraft_v4_1` (clean product mockup) — or your own photo|The “real world” reveal must read photographic | |Voiceover |`higgsfield_audio_generate` (voiceover / `seed_audio`) or `text2speech_v2` (ElevenLabs / MiniMax / Seed Speech presets)|Deadpan conversational delivery; can voice-clone from a sample | | |Music bed |`higgsfield_audio_generate` seed_audio prompt (or game audio `sonilo_music`) |“Lo-fi synthwave, ~95–100 BPM, keeps under the voice” | | |Captions burned in |`video-subtitler` skill |— |Timed phrase-chunks, white bold, centered low-third, drop shadow | |Assembly / trim / mix |`video-montage` skill (+ Higgsedit for cut-on-frame precision) |— |Concatenate clips, trim generated 4s+ takes to exact 2s/1s shots, music under VO | |Prompt text discipline |`seedance-cinematic-prompter` skill |— |Owns blocking/optics/lighting/first-frame wording for each shot | Constraint that shapes everything: **video models generate 4s minimum** (Seedance 2.5: 4–30s integer; MiniMax H3: 4–15s). The source’s 2s and 1s shots cannot be generated as-is — generate 4–6s coverage of each beat and **trim to the exact cut in assembly**. The cut point lives in the edit, not the generation. **5. Shot-by-shot reconstruction workflow (apply to your character/env/dialogue)** **Stage 0 — Script grid (write yours, keep the beat DNA)** Map your monologue into the same six-function structure. Punchline-first logic: 1. Hook (2s) — meme proverb, three captions, one hand gesture per caption. 2. Context (9s) — the absurd situation, told with open-hand gestures and shrugs, no prop. 3. Anecdote (7s) — one story; introduce a small prop; smirk on the punchline; one *scoff* beat. 4. Payoff (4s) — the reveal: source of the “goal” (in the source: the shirt graphic). Chest beat + point. 5. CTA (1s) — one short line, deadpan stare. 6. Product (still) — URL + “link in bio”. **Stage 1 — Assets (all image generations, then register element)** - Character sheet: original design, three-quarter + face; include a wearable graphic the payoff shot can reveal. Register as a persistent element. - Environment keyframe: your location (it does not need to be a balcony — it needs *height +* *distance + haze* to give the rim-light and depth the grammar uses). - Product/merch photo: shoot or generate the “real” version. **Stage 2 — Audio first (recommended order)** - Record/generate the VO with deadpan pacing and the pause beats; keep lines short (2–6 words). - Generate the lo-fi bed at ~95–100 BPM, mix under the VO. - Use the VO as an **audio reference** in generation so lip-sync matches the scripted timing (role: "audio"). This is what makes the meme cadence land. **Stage 3 — Per-shot generation** For every shot keep a **lighting lock line** identical across prompts (e.g. “warm golden-hour sun from camera-left back, strong white rim light, sea-of-clouds haze, 4:3”) and the element reference. Prompt skeleton per shot (fill with your own): - **Shot 1** (Seedance 2.5, ~5s coverage): MCU, eye-level, subject screen-right leaning on [railing], slow push-in, direct eye contact; scripted gesture: open palm on first phrase, clench fist on second, pull to chest on third; captions appear with each phrase. - **Shot 2** (MiniMax H3, ~9s): WS full body, camera back, slight low angle, static with slow handheld sway; weight shifts, open-hand gestures per clause, shrug, head shake on the “nothing happened” beat; background activity: clouds drift, distant lights blink. - **Shot 3** (Seedance 2.5, ~7s): MCU tighter 3/4 angle, slow lateral drift right; [prop] in hand; smug smirk, head tilt, brow raise on punchline. - **Shot 4** (Seedance 2.5, ~5s coverage): MS waist-up, slow push-in; beat chest with fist, double thumbs, point at the graphic; graphic readable; shallow DOF background. - **Shot 5** (Seedance 2.5, ~4s coverage): CU face, locked-off, deadpan stare, chin slightly down; deliver CTA line. - **Shot 6**: static product photo — photographic, white studio, garment on model. **Stage 4 — Assembly** - Trim each take to its exact timestamp; cut on punch words/caption changes. - Concatenate in order; place product still; mix VO (dominant) + music (bed) + one accent SFX if any. - Burn captions: white bold sans, centered low third, soft shadow, **chunks synced to phrases** — the caption change is the cut trigger. **Stage 5 — QC checklist** - [ ] Same character proportions/wardrobe across all five 3D shots (element held, no morphing). - [ ] Rim light direction identical in every shot. - [ ] Shirt graphic in shot 4 matches the real product in shot 6. - [ ] Lip-sync matches the VO on every line; gestures land on the same words. - [ ] Cut-to-caption sync: every cut sits on a punch word or caption change. - [ ] No empty seconds: VO and music run continuously. - [ ] 4:3 framing applied (or letterboxed) — no accidental 9:16.