create-photo-grid-ad

name	create-photo-grid-ad
description	Produce a Photo-Grid promo card — a silent, music-optional feed-placement video where a scrolling 2×N tile grid (product photos + short brand video clips + big-type offer tiles + a promo-code tile) animates under a centered wordmark, headline, sub, and outline feature chips. 9:16 or 4:5, ~10s, $0 HTML-hyperframe pipeline (no T2V, no VO). Recreates the Meta "photo-grid mosaic" template (e.g. "Our New Band 2.0 / LAUNCH30"). Validated on Hume Health (Father's Day promo, 2026-06-03). Use when the brand wants a clean editorial feed card that works muted; NOT for kinetic-BG vignette (use create-vignette-ad), UGC, or VO-driven ads.
status	ACTIVE — validated on Hume Health (pick-05 Father's Day promo card, 2026-06-03)

create-photo-grid-ad

Photo-Grid format = a continuously-scrolling tile grid (product photos + short brand video clips + big-type offer + promo-code tiles) under a wordmark + headline + chips. Silent, music-optional, sub-12s, loopable. A single HTML hyperframe rendered frame-by-frame via Playwright + ffmpeg — $0 spend, ~30s compute. This is the cheap, editorial sibling of create-vignette-ad (which lays product cutouts over a paid T2V kinetic background).

Purpose

Produce a short-form (8–12s) vertical or 4:5 Photo-Grid promo card for a brand by:

1. Gathering brand assets (clean product PNGs, logo SVG, brand palette, and short clips harvested from the brand's own existing ad corpus — HD source preferred over Pexels).
2. Authoring ONE HTML hyperframe: centered wordmark, big headline + sub, a 2×N tile grid that scrolls horizontally for the whole duration behind an edge-fade mask, and outline feature chips.
3. Filling the grid with a rhythm of tile kinds — product photos, short looping brand video clips, big-type offer tiles (e.g. "25%" / "OFF" in a display serif), and one dark promo-code tile.
4. Rendering it frame-stepped (Playwright screenshot per frame, videos played in real time) and encoding to MP4 with ffmpeg.
5. Optionally laying an instrumental music bed under it (EBU R128 loudness-normalized) — the base master is silent and works muted.

Use when:

• The brand has clean product photography AND/OR an existing ad/video corpus to harvest tiles from.
• The aesthetic target is an editorial, premium feed card that reads instantly while muted.
• A promo/offer (discount + code) or a seasonal hook ("For dad, by Hume.") drives the card.
• Budget is ~$0 and turnaround is minutes.

Do NOT use for:

• Kinetic-background atmospheric ads with product cutouts → use create-vignette-ad (paid T2V BG).
• UGC creator / talking-head ads → use skills/molecules/ugc-ad/*.
• VO-driven explainers or demos → use a VO-anchored beat-video molecule.
• Long-form (>15s) → the format's read breaks past ~12s.

Inputs

Required:

• brand_url — e.g. https://humehealth.com (for product PNGs, logo, palette).
• headline — 2–4 words, 1–2 lines, e.g. "For dad, by Hume.".
• offer — discount + code, e.g. 25% OFF + DAD25 (or a non-promo hook if the card is brand-led).

Optional (with autonomous defaults):

• aspect_ratio — 9:16 (default, 1080×1920) or 4:5 (1080×1350).
• duration_s — default 10.0 (range 8–12).
• sub — one line under the headline, e.g. "The longevity band, gift-wrapped.".
• grid — rows × tiles-per-row (default 2×5) + tile size (default 360×380, 24px gap).
• chips — 0–3 outline feature pills, e.g. ["14 DAY BATTERY", "NEW PATENTED STRAP"].
• palette — { ink, ink_soft, bg, bg_tile, bg_band, accent } hex; default derived from the brand homepage CSS (top frequencies) + the brand accent.
• fonts — one display sans + one display serif (default Cabinet Grotesk + Zodiak via Fontshare).
• music_brief — free-text; when set, an instrumental bed is generated and muxed. Default: silent.
• source_clips — brand ad-corpus clips to extract video tiles from (default: scan the brand's existing-ads corpus for sustained-shot HD moments; fall back to Pexels portrait stock).

Prerequisites:

• Playwright Chromium + ffmpeg/ffprobe on the machine.
• librsvg (rsvg-convert) if the logo is an SVG that needs rasterizing.
• PEXELS_API_KEY (optional) only if the brand has no usable corpus and stock tiles are needed.
• For a music bed: the Gooseworks ElevenLabs proxy (bills the user's credits) or a stored asset.

Composed Atoms

• source-brand-assets — scrape product PNGs from the brand Shopify/CDN, brand palette hex from the homepage CSS, and the logo SVG from the footer.
• create-motion-graphics-hyperframes — author + frame-step render the HTML hyperframe (wordmark + headline + sub + scrolling grid + chips) via Playwright; the renderer scaffold (initRenderer, tw/easeOut/clamp01) drives per-frame opacity/transform.
• stitch-videos-ffmpeg — extract sustained-shot tile clips from the brand corpus (scale→crop→scale for low-bitrate sources), apply setpts slowdown so no tile loops within the window, and encode the frame sequence to MP4.
• create-music-elevenlabs — optional instrumental bed via the ElevenLabs music endpoint (generate ~2s longer than the card, trim).
• add-music-track — EBU R128 loudness-normalize the bed (loudnorm I=-18:TP=-2:LRA=9) and mux it as a SEPARATE ffmpeg pass with explicit -map 0:v:0 -map 1:a:0.
• watch — /watch:watch for end-to-end QC on the rendered master.

Workflow

Phase 0 — Brand asset gathering (autonomous)

1. Scrape clean product PNGs from the brand PDP (highest-res available). Promote the best PDP shots into the shared brand-assets/reference-photos/ namespace so later cards reuse them.
2. Scrape brand palette + logo — top hex frequencies from the homepage CSS; logo SVG from the footer/theme assets. Derive { ink, ink_soft, bg, bg_tile, bg_band, accent }.
3. Harvest tile video clips — scan the brand's existing-ads corpus for HD, sustained-shot moments (product macro, app screen-record, a clean lifestyle beat). Probe existing-ads/*.mp4 for resolution BEFORE reaching for any meta-library/*.mp4 — meta-library is 360p delivery and reads blurry upscaled; existing-ads is the HD pool. Avoid Pexels when the brand has its own corpus.
4. Audit clip captions for compliance — brand ad b-roll sometimes carries on-screen claims (e.g. health-risk overlays). Skip any clip whose caption makes a sensitive claim.

Output: brief.md with the chosen tiles (product photos + clip windows), palette swatches, headline/sub/offer/chips copy, and aspect/duration.

Phase 1 — Author the hyperframe (autonomous)

1. Author one HTML file at the target canvas (1080×1920 for 9:16). Load the display sans + serif from Fontshare. Lay out: centered wordmark (top), headline (big extrabold, ≤2 lines), sub, a grid-viewport with an edge-fade mask-image, a grid-track of repeat(tiles-per-row, W) columns × N rows, and outline chips near the bottom.
2. Fill the tiles with a rhythm — product-photo tiles (object-fit: contain on a band background), video-clip tiles (<video autoplay muted loop playsinline>, object-fit: cover), big-type offer tiles (display serif, e.g. "25%" / "OFF"), and one dark promo-code tile (label + value with an accent-colored prefix, e.g. DAD in accent + 25 in white).
3. Drive motion in initRenderer(duration, t => …): wordmark fade-in (0–0.4s), headline drop+fade (0.1–0.7s), sub fade (0.5–0.9s), then a continuous linear translateX scroll of the track across its full overflow for the whole duration (with a soft eased entry), and chips fade-up (1.5–1.9s).
4. All asset paths must be RELATIVE to the HTML file — Playwright renders blank for file:// absolute paths.

Phase 2 — Render + optional music (autonomous)

1. Frame-step render: launch Chromium with --autoplay-policy=no-user-gesture-required, wait for fonts + every <video>'s canplay + play(), then for each of duration×fps frames call renderAt(t), double-requestAnimationFrame, screenshot, and sleep to keep real-time video playback. Encode the PNG sequence to MP4 (h264, yuv420p, +faststart, 25fps) → silent master.
2. Tile clip prep: for each corpus clip, setpts=N*PTS where N = duration / clip_len so the tile never loops within the window (2–4× slowdown also reads as premium slow-mo).
3. Optional music: when music_brief is set, generate an instrumental bed, loudness-normalize, and mux as a separate pass (-map 0:v:0 -map 1:a:0 -shortest). Verify audio bit_rate ≈ 192k.
4. Self-QC via /watch:watch: scroll is smooth and continuous; headline + chips legible; type tiles don't bleed; no video tile visibly loops; promo code readable; duration ≈ target.

Decision Rules

1. One HTML, one render script — skip the full beat-video pipeline. When the brief is a static/ scrolling photo grid or feed promo card, do NOT bring in VO, Whisper, beat manifests, concat, or captions. Author one hyperframe + one render pass. (~30s compute, $0.)
2. Grounded brand corpus beats Pexels for tile content. The brand's own ads carry authentic color grade, lighting, and product fidelity stock can't match. Reach for Pexels only when the brand has no usable corpus.
3. HD source only — probe existing-ads/ before meta-library/. Meta-library ads are 360p delivery; upscaled into a 1080-wide tile they read obviously stock-y.
4. grid-template-columns: repeat(N, W) is tiles-PER-ROW, not total tiles. Setting it to the total tile count packs everything into row 1 and leaves later rows empty.
5. setpts slowdown, never <video loop> with short clips. One visible loop-restart makes the whole tile read as cheap stock. Slow each clip to fill the window (or ping-pong).
6. Edge-fade the grid viewport with a horizontal mask-image gradient so scroll tiles appear/ disappear softly instead of popping at the edges.
7. Pair one display sans with one display serif when the card mixes photo tiles and big-type tiles — the serif (offer/price) adds rhythm and an editorial "sales-print" weight.
8. Serif display digits run ~20–30% wider than geometric sans — budget type-tile width at ~65% of nominal and re-check fit after any aspect change.
9. Aspect change cascades into every type size. Converting 4:5 ↔ 9:16 isn't free real estate — re-audit headline, tile, type-in-tile, and chip sizes against the new bounds.
10. The base master is SILENT and must work muted. Music is an optional bed, never load-bearing — feed placements autoplay muted. When a bed is used, keep it instrumental and loudness-normalized.
11. ffmpeg mux: explicit -map 0:v:0 -map 1:a:0 + a SEPARATE pass from the frame encode. Default mapping silently produces ~1 kbps garbage audio.
12. Style tokens pass a taste gate — website palette/type is a candidate, not law. Background must clear a legibility gate vs the headline/offer ink (≥ 4.5:1); when the brand color fails, use near-white (#F4F8FB/white) with dark ink, or a dark brand-neutral with light ink. Use the brand font ONLY if it's a loadable webface (Google Fonts/Fontshare); else the house faces (Cabinet Grotesk + Zodiak). Never author type below the muted-feed floors (1080-wide 9:16): headline ≥ 132px, sub ≥ 32px, offer/price ≥ 140px, code value ≥ 88px, code label ≥ 22px, chips ≥ 26px. (pick-05 ships / and 154 / 38 / 156 / 104 / 26 / 30.)

Output

<project>/
├── brief.md
├── hyperframes/
│   ├── hyperframe.html          # the card layout + initRenderer motion
│   └── _shared.js               # initRenderer / tw / easeOut / clamp01 scaffold
├── assets/
│   ├── grounded/                # HD tile clips extracted from the brand corpus
│   └── music/                   # optional bed (raw + loudnorm)
├── render.py                    # Playwright frame-step + ffmpeg encode
└── finals/
    └── master-<aspect>.mp4      # silent master (+ -music-<variant>.mp4 when a bed is muxed)

The shipped master is the silent master-<aspect>.mp4; music variants are siblings.

Quality Checks

• Duration within 8–12s; aspect matches requested (9:16 = 1080×1920, 4:5 = 1080×1350).
• Wordmark, headline, sub render in loadable fonts at the right sizes (no clipping/bleed).
• Type sizes ≥ the muted-feed floors (headline ≥132 / sub ≥32 / offer ≥140 / code ≥88 / chip ≥26 px at 1080w).
• Background passes legibility vs the ink (≥4.5:1); not a low-contrast website-lifted combo.
• Photo tiles are real product photos (≥720px short-edge, no favicons/logos, no >1.5× upscale).
• Grid scrolls continuously across its full overflow for the whole duration; entry is soft.
• Grid viewport has the edge-fade mask (tiles don't pop in/out at the edges).
• Every video tile is HD and never visibly loops within the window (setpts applied).
• Type tiles ("25%", "OFF") fit inside the tile with margin; serif sizing re-checked for aspect.
• Promo-code tile reads as a voucher (label small, value huge, accent prefix).
• Feature chips legible as outline pills; no compliance-sensitive claim copy survives any tile.
• Base master plays cleanly muted in QuickTime + browser; if a bed exists, audio ≈ 192k AAC, −18 LUFS.

Failure Modes

Failure	Cause	Recovery
All tiles land in grid row 1, later rows empty	grid-template-columns set to total tile count	Use repeat(tiles-per-row, W)
Tile video visibly restarts mid-card	short <video loop> clip not slowed	setpts=N*PTS (N = duration / clip_len) or ping-pong
Tiles read blurry/pixelated	extracted from 360p meta-library/ source	Re-extract from HD existing-ads/ corpus (probe resolution first)
Playwright renders a blank card	file:// absolute asset paths	Make every asset path relative to the HTML file
Type tile text bleeds past the tile edge	serif display digits wider than expected	Drop size ~20–30% + tighten tracking; budget width at 65% nominal
Entering/leaving tiles pop at grid edges	no viewport mask	Add mask-image: linear-gradient(90deg, transparent, #000 4%, #000 96%, transparent)
Bottom of a 9:16 card looks empty	grid too short for the taller canvas	Add a 3rd tile row, grow tiles, or push chips up (or keep the whitespace as "premium minimal")
Final MP4 audio is ~1 kbps	ffmpeg mux without explicit map	-map 0:v:0 -map 1:a:0 as a separate pass from the frame encode
Clip caption makes a health/risk claim	brand b-roll overlay copy	Drop that clip; pick a caption-clean sustained shot

Tests

See tests/:

• smoke-test.md — structural pass (one hyperframe → silent MP4 at target aspect/duration).
• sample-input.md — Hume Health Father's Day promo card worked example.
• expected-output.md — what a successful render produces.
• human-test.md — manual acceptance criteria (watch the card muted).
• verifier.md — ffmpeg/ffprobe assertions.

Related

• Reference build: Hume Health pick-05 Father's Day promo card — HOW_TO_MAKE_THIS_VIDEO.md + LEARNINGS.md (18 load-bearing learnings; $0, 9:16, 10-tile scroll with HD corpus video tiles).
• Remix wrapper: remix-photo-grid-ad-from-sample — the app's Photo-Grid format tab calls this to remix a published card for a new brand.
• Sibling format: create-vignette-ad — product cutouts over a paid T2V kinetic background (the atmospheric, not editorial, silent format).
• Registry entry: formats.json → "photo-grid".