--- title: "E-commerce Lookbook — cinematic fashion b-roll + a pixel-perfect product-card overlay" tier: hero format: short-form-video theme: commerce | fashion | ecommerce | product persona: e-commerce brand, fashion creator, DTC marketer, social shopping duration: "1 product → 30–45s vertical lookbook in ~6–12 min" budget_usd: "$0.80–$4.00 per reel (5–7 beat keyframes + Ken-Burns/motion + TTS + music + hyperframes overlay)" caps: ["generate_project", "ltx-q-i2v", "seedance-i2v", "chatterbox-tts", "music", "ffmpeg-concat", "ffmpeg-mux", "hyperframes-render"] skills: ["short-form-video", "hyperframes-overlays"] showcases: ["/chapters/ecommerce-lookbook-example.html"] status: "live (2026-06-09) — rendered E2E. The example reel ('The Camel Trench') was produced exactly via this recipe (Ken-Burns base + a product-card hyperframes-render overlay)." reliability: 4.4 # Ken-Burns default path: keyframes −0.3, tts+music −0.4, 7 stages −0.4, +0.5 proven E2E showcase; the i2v upgrade path swaps Ken-Burns for video (−0.6) ≈ 3.8 --- # E-commerce Lookbook — cinematic fashion b-roll + a pixel-perfect product card Point this at a product — a garment, an accessory, a hero SKU — and it builds a **shoppable vertical lookbook**: cinematic AI fashion b-roll for the *feel*, and a single timed-HTML overlay that *is* a real product card. The b-roll carries the styling and the mood; the overlay carries the commerce UI — brand name, a free-shipping badge, the product name + **$189 price**, a ★★★★★ rating, a size selector, an "only 4 left" inventory bar, and a **BUY NOW** button. AI video can never render a legible price or a crisp size selector — and you must **never** trust a model to render a price — so this playbook keeps all of that in the HTML layer. ## What you'll get > **A finished reel = vertical 9:16 + voiceover + music bed + a pixel-perfect product-card HTML overlay + 30–45s.** Pass `aspect_ratio: "9:16"` to `generate_project` (it threads to every scene; keep the garment consistent across shots with `character_anchor:"always"` + a restated wardrobe — the model often returns 4:3, so the Ken-Burns step cover-crops to exactly 1080×1920, see the note below), animate in waves of ≤4–5 with `resolution:"auto"` (or use the Ken-Burns stills fallback when fal's video tier is flaky), stitch with `create_media model_override:"ffmpeg-concat"`, then lay `chatterbox-tts` narration over a ducked `music` bed with `create_media model_override:"ffmpeg-mux"` + `audio_fill:"loop"` (so audio covers the FULL reel — no silent tail), then composite a single `hyperframes-render` document — the product card — over the whole reel. A silent montage with the price melted into a blurry frame is a draft, not a deliverable — see the `short-form-video` skill's *definition of done* and the `hyperframes-overlays` skill for the overlay pattern. A 30–45s **vertical (9:16, 1080×1920)** lookbook: a hook → styling beats → detail beat → call-to-action script, 5–7 cinematic beat keyframes of the *same garment*, slow motion (i2v or Ken-Burns), a warm brand VO over a chic bed, **and** a pixel-perfect HTML product card — brand, badge, name, price, rating, size selector, inventory bar, and BUY NOW — that animates in over the reel. ## How the agent should run this (interaction contract) 1. **CONFIRM** — one line: product + price/sizes/rating + the choices that matter (motion path: **Ken-Burns default** vs i2v upgrade; 5–7 beats; voice tone) + "~$0.80–$4.00, ~6–12 min". Commerce data (price, sizes, stock) MUST come from the user — never invent it; that's the one question worth asking. 2. **PREVIEW CHECKPOINT** — the **keyframe contact sheet** (5–7 stills, ~$0.40) **plus the product-card HTML shown as a static frame**. Both before any motion/TTS/music spend — garment drift or a wrong price gets fixed here for cents, not after the mux. 3. **NARRATE** — one line per stage with job_id + ETA ("animating beat 3/6, ~90s each"); poll `get_create_media` every 10–15s; i2v in waves of ≤4–5; flag the hyperframes render ("compositing overlay, ~1–2 min"). 4. **FAIL GRACEFULLY** — keyframe 422/timeout: soften prompt → `flux-dev` / `krea-2-large`. i2v hang: `seedance-i2v` → `ltx-q-i2v resolution:"auto"` → **Ken-Burns** (the live showcase shipped on Ken-Burns — state the swap in one line, don't apologize). TTS or music fail → ship VO-only or bed-only and say so. ≤2 retries per clip; partial reel beats nothing. 5. **DELIVER** — final MP4 + one honest note ("commerce UI is pixel-perfect HTML; garment held on N/N beats") + ONE next step: swap the detail beat to a `kontext-edit` of the real product photo. ## The cap chain (the PROVEN recipe) ``` product (the garment / SKU — and its real price, sizes, rating) → agent writes the script: hook → styling beats → detail → call-to-action (this is BOTH the brand VO AND the on-screen card copy) → generate_project # one cinematic 9:16 fashion keyframe per beat aspect_ratio: "9:16" character_anchor: "always" # KEEP THE GARMENT CONSISTENT across shots restate the wardrobe in every scene prompt (the camel trench held this way) image: gpt-image | flux-dev | recraft-v4 | krea-2-large → motion (pick ONE — Ken-Burns is the DEFAULT, i2v is the upgrade): (a) Ken-Burns DEFAULT # ffmpeg zoompan slow-zoom on each 9:16 still → ~7s clips # deterministic, exactly 9:16, zero garment drift. The example reel used THIS. (b) i2v upgrade # create_media action:"animate" — ltx-q-i2v / seedance-i2v, resolution:"auto", # slow cinematic push / fabric drift — or kling-i2v / pixverse-i2v / cosmos-3-i2v. # Worth it for fabric motion; costs reliability (fal video tier is flaky). → create_media action:"tts" model_override:"chatterbox-tts" # warm brand-voice narrator → create_media action:"music" lyrics_prompt:"[Instrumental]" # chic instrumental bed ("downtempo fashion-runway, no vocals") → create_media model_override:"ffmpeg-concat" # stitch the clips → base reel → create_media model_override:"ffmpeg-mux" audio_fill:"loop" # VO over a ducked music bed, covering the FULL reel → hyperframes-render # THE PRODUCT-CARD LAYER — one timed HTML doc over the whole base reel # ≤15s reel → create_media model_override:"hyperframes-render" # >15s reel → call the cap directly (tool /run/hyperframes-render # or raw SDK /inference) with duration:N ``` ## The differentiator — the product-card overlay HTML `hyperframes-render` runs an agent-generated HTML document in Chromium and composites it over the base reel for `duration` seconds. The whole layer is **one** HTML file driven by CSS `@keyframes`. It doesn't just caption — it reconstructs a real e-commerce product card, with exactly the data a shopper acts on: - the **brand** ("Maison Atelier") + a **"Free shipping"** badge, - the **product name + price** ("The Camel Trench · **$189**") — the number that must be exact, - a **★★★★★ 4.9 (320)** rating row, - a **size selector** (XS S [M] L XL, M highlighted), - an **"Only 4 left in size M"** inventory bar (scarcity), - a **BUY NOW** button. ```html
FREE SHIPPING
Maison Atelier
The Camel Trench
$189
★★★★★ 4.9 (320)
XS
S
M
L
XL
Only 4 left in size M
BUY NOW
``` **The 15s cap split:** a 35s reel exceeds `create_media`'s 15s `duration` cap. For long reels, call the capability directly — the tool endpoint `/run/hyperframes-render` or raw SDK `/inference` — with `duration: 35`. Reels ≤15s can go through `create_media model_override: "hyperframes-render"`. ## Script structure (beats) 1. **Hook (0–4s)** — the desire, stated cold. "The one coat that finishes every outfit." 2. **Styling beats (2–3, 1 beat each)** — the garment in context: street, café, evening. Same garment, different scene. 3. **Detail beat** — the close-up that sells the make: lapel, stitch, fabric drift. 4. **Call-to-action (last beat)** — the card fully revealed; "Tap to shop The Camel Trench." ## Quality bar - **Native 9:16** — every keyframe and every clip is 1080×1920, NOT post-cropped from 16:9. - **Garment consistency** — the *same* coat across every shot, held by `character_anchor:"always"` + a restated wardrobe in each prompt. The example reel's camel trench stayed consistent this way. - **Deterministic commerce UI** — brand, price, rating, sizes, inventory, and CTA live in the `hyperframes-render` layer. **Never** trust the model to render a price. - **VO + music** — `chatterbox-tts` brand voice over a ducked chic bed; silent is a draft. - **30–45s** — long enough to style the look and land the CTA, short enough to finish. ## What can disappoint (cap ceilings) i2v can shift the garment's cut, colour grade, or drape **mid-clip** even when the keyframe was perfect — i2v style stability is not a promise any cap makes; Ken-Burns is immune. And no image/video model ever renders a readable price — that lives in HTML, always. ## Watch-outs - **Never put the price (or any commerce data) in the image.** This is the whole point of the playbook. AI text rendering is the hardest known failure mode — a melted "$189" or a garbled size chip destroys trust and conversion instantly. Render price, rating, sizes, and inventory in the HTML layer where they're pixel-perfect every time. - **Garment drift is the real risk — anchor it.** Any text-to-image model has zero cross-shot consistency by default. Use `character_anchor:"always"`, restate the wardrobe ("a camel double-breasted trench, belted") in every scene prompt, and prefer a reference-image edit (`kontext-edit`) for the detail beat. The example reel's trench held because the wardrobe was restated and anchored. - **fal video tier is flaky — mitigate, don't fight it.** The video tier has real latency. Mitigations, in order: (1) prompt **image-led** fashion subjects; (2) try alternate video caps (`kling-i2v`, `pixverse-i2v`, `cosmos-3-i2v`, `ltx-q-i2v`); (3) render in **waves of ≤4–5** and poll each to *done*; (4) fall back to the **Ken-Burns stills** path (ffmpeg `zoompan` slow-zoom on each 9:16 keyframe → ~7s clips). The example reel honestly used the Ken-Burns fallback — reliable and exactly 9:16. - **4:3 keyframes → cover-crop to 1080×1920.** The image model often returns 4:3 even with `aspect_ratio:"9:16"`. The Ken-Burns step's `zoompan`/`scale`+`crop` cover-crops to exactly 1080×1920 — so the base reel is true vertical regardless of what the model returned. - **`hyperframes-render` runs agent-generated HTML in Chromium with `file://` access.** Only ever feed it **agent-authored** HTML — never paste untrusted third-party markup. Treat the overlay document as code you wrote. ## Cross-surface (MCP / CLI / Webapp) - **MCP:** `generate_project` (9:16 beats, `character_anchor:"always"`) → `create_media` (animate `resolution:"auto"`, TTS, music) → `ffmpeg-concat` → `ffmpeg-mux` (`audio_fill:"loop"`) → `hyperframes-render` overlay (≤15s via `create_media model_override`, >15s via the cap directly with `duration:N`). - **CLI:** `livepeer story " lookbook, 9:16, same garment across shots"` → TTS + music via `livepeer media create` → `livepeer media export` to stitch → call `hyperframes-render` for the product-card overlay (long reels via raw `/inference duration:N`). Ken-Burns fallback via `ffmpeg zoompan` if the video tier rejects. - **Webapp chat:** `make an e-commerce lookbook for with a product-card overlay (price, sizes, BUY NOW)` — the agent drafts the hook→styling→detail→CTA script, renders 9:16 b-roll with the garment anchored, and offers the product-card finishing as a follow-up. ## Execution mode (autopilot / director / auto) This is a multi-scene project, so pass a `mode` to `generate_project` to control the **pre-finish gate** (after every scene renders, before delivery): `autopilot` (certain brief → ship in one call, no pause), `director` (always pause for a human pick/steer before delivery), or `auto` (recommended — ship UNLESS the auto-critic is unsure: score < `confidence_threshold` (default 0.8), verdict `iterate`, or a flagged shot). Setting `mode` runs the project async (returns a `job_id`); resume a director/auto pause with `review_checkpoint` + `resume_from_checkpoint` (decision `continue` to deliver / `abort` to discard). Omit `mode` for legacy straight-through behavior.