create-vignette-ad

create-vignette-ad

Vignette format = product cutouts laid over a kinetic background video. Music-led, sub-12s, loopable. The inverse of video inlays. Validated 6-variant comparison run with 3 sub-archetypes (V-PURE / V-TILE / V-CARD). This molecule focuses on V-CARD (cold-open card → product carousel → annotated end card).

Purpose

Produce a short-form (9-12s) vertical or square Vignette ad for a brand by:

1. Scraping brand assets (products, logo, palette, IG B-roll for inspiration)
2. Defining 1-2 BG concepts grounded in brand identity (NOT generic "abstract macro")
3. Sourcing BG via Pexels stock FIRST (free, native 9:16 candidates: ink in water, sheer fabric wind, liquid bloom) — only fall back to T2V (Veo/Kling/Seedance) if Pexels doesn't deliver. Pexels-substitution validated on Clinikally; saves $7+ per run.
4. Compositing per-variant: BG (palette-aware dim — heavier on light BG, lighter on dark) + product cutouts (height-anchored at vertical-center for mixed-shape SKUs, width-anchored for similar-aspect SKUs) + cold-open card + annotated end card (WHITE logo variant for dark BGs)
5. Adding ambient instrumental music bed + EBU R128 loudness norm
6. Delivering N variants for selection (N = num_concepts when BG=Pexels; N = num_concepts × num_models when BG=T2V)

Use when:

• Brand has clean product photography (Shopify PDP shots ideal — birefnet strips cleanly)
• Brand has an active IG presence (gooseworks scrape gives brand-coded reference)
• Aesthetic target is music-led atmospheric, NOT UGC creator or talking-head
• Budget allows multi-variant comparison ($7-10 per run for 6 variants)

Do NOT use for:

• UGC ads with creator → use skills/molecules/ugc-ad/*
• VO-driven explainer → use skills/molecules/explainer-video/*
• Single static product hero shot → use skills/atoms/product-images/*
• Long-form (>15s) → format breaks past 12s

Inputs

Required:

• brand_url — e.g. https://motherscience.com
• products[] — 1–3 product handles/slugs. 1 = single-cutout focus, 2-3 = carousel.

Optional (with autonomous defaults):

• aspect_ratio — 9:16 (default, mobile) or 1:1
• duration_s — default 10.5 (range 9-12)
• sub_archetype — V-CARD (default — cold-open card + carousel + endcard), V-PURE (pure carousel no text), V-TILE (persistent social-tile layout — heavier production)
• bg_concept — free-text concept brief. If omitted, derived from brand category (skincare → liquid macro / molecular; cleaning → foam; supplement → ingredient bloom)
• bg_model_set — ["veo-3-1-preview","kling-v2.5-turbo-pro","seedance-2.0"] (default — fire all 3 for comparison) or subset
• bg_concepts_count — 2 (default — fire 2 concept variants for choice) or 1 (cheaper, less optionality)
• cold_open_text — default 3 lines, e.g. "100% / PROVEN / RESULTS". 3-5 words max.
• end_card_lines — small print annotations. Defaults: ["EST {year}", "{ingredient}", "NOVEL MOLECULE", "{claim}"]
• music_brief — free-text. Default matched to brand category (skincare → ambient minimal electronic 90 BPM)

Prerequisites:

• FAL_API_KEY in env (.env root) — for birefnet, Kling, Seedance, ElevenLabs
• PEXELS_API_KEY (optional, in gtm-goose/.env) — for fallback stock
• Higgsfield CLI authed at ~/.config/higgsfield/credentials.json — for Veo 3.1
• npx gooseworks credits returns balance — for IG scraping
• Homebrew installed: ffmpeg, librsvg

Composed Atoms

• source-brand-assets — WebFetch product PNGs from Shopify, brand color hex from homepage CSS, brand logo SVG from footer
• remove-backgrounds — strip product PNG BG to clean alpha via fal-ai/birefnet/v2 (auto-handles overlay badges + drop shadows)
• create-video-veo-3-1-higgsfield — generate cinematic BG variant via Higgsfield CLI (Veo 3.1 ultra preview, 8s, 16:9 or 9:16)
• create-video-kling-v2-5-turbo-pro-fal — generate BG variant via FAL Kling v2.5 Turbo Pro T2V (10s, 1:1 supported)
• create-video-seedance-2-fal — generate BG variant via FAL Seedance 2.0 T2V (8s 720p, fastest queue)
• create-motion-graphics-hyperframes — render cold-open card + annotated end card via PIL + rsvg-convert for brand SVG logo (alternative to Playwright)
• create-music-elevenlabs — instrumental music bed via fal-ai/elevenlabs/music endpoint, generate 12s, trim to target
• composite-split-screen — used as ffmpeg filter_complex pattern reference for multi-layer composites with per-layer enable + scaling
• stitch-videos-ffmpeg — used for the audio mux pass per memory feedback_ffmpeg_lcut_endcard_recipe.md (separate pass from composite)
• add-music-track — apply EBU R128 loudness normalization via acompressor + loudnorm I=-18:TP=-2:LRA=9 before mux
• watch — /watch:watch external skill for end-to-end QC on each final variant
• External: gooseworks scrapecreators /v1/instagram/profile — scrape brand IG reels for B-roll inspiration (NOT used as BG directly — only for visual language reference)

Do NOT use: Klap, captions, Whisper, Soul ID, UGC-creator atoms — Vignette is silent-text only.

Workflow

Phase 0 — Brand asset gathering (autonomous)

1. Scrape product PNGs from brand Shopify (or other CDN):

   curl -sL "${brand_url}/products/${handle}.json" | jq -r '.product.images[0].src'
   

   Download highest-res; PDP shots typically 1500-2000px on white BG.

2. Background-remove via fal-ai/birefnet/v2 ($0.02 each). Birefnet auto-handles overlay badges (Glamour seals etc.) AND drop shadows. Validate alpha: ≥20% transparent, ≤8% partial-edge. NB2-edit retouch only if validation fails.

3. Scrape brand color hex from homepage — count #xxxxxx patterns, take top 5 frequencies. Logo SVG via footer link or Shopify theme assets path.

4. Scrape brand IG reels via gooseworks (~$0.02):

   npx gooseworks call scrapecreators /v1/instagram/profile \
     --query='{"handle":"<brand_ig_handle>"}'
   

   Parse out video URLs. Download 3-4 reels. Do NOT use brand B-roll as the BG (felt like remix when validated) — use as REFERENCE for visual language + palette + concept inspiration.

5. Define BG concept(s) grounded in brand identity:

   • NOT generic "abstract macro" — that's a style not a theme. Pexels returns stock 3D art for those queries.
   • DO brand-category specific: skincare → ingredient bloom / molecular structure / serum drop; supplement → liquid swirl / pour; cleaning → foam macro
   • Mother Science default: (α) Chrome Molecular Swarm, (β) Black Ink in Cream Bloom

Output: <project>/brief.md with cutout previews, brand palette swatches, IG reel thumbnails, BG concept(s) + rationale, estimated spend.

🚦 HUMAN GATE 1 — Brief approval

Show user: brief.md + sub-archetype recommendation + BG concept proposals + estimated spend ($7-10 typical for 6-variant comparison).

User approves OR edits brief.md.

Phase 1 — BG generation (post-gate, autonomous)

1. Write per-model T2V prompts per concept (see references/bg-video-sourcing-rules.md). Key rules:

   • Use PHASE markers (PHASE 1 (0-2s):), NEVER "Shot N:" labels (force hard cuts)
   • Explicitly say "NOT slow motion" + add to negative prompt — Vignette format wants kinetic real-time motion
   • Lock camera ("camera locked tripod, no movement")
   • Explicitly prohibit text, products, hands, people in negative prompt

2. Fire all variants in parallel via ThreadPoolExecutor:

   • Veo 3.1 via Higgsfield CLI: higgsfield generate create veo3_1 --prompt "..." --aspect_ratio 9:16 --duration 8 --model veo-3-1-preview --quality ultra --wait --json
   • Kling v2.5 Turbo Pro via FAL: fal_client.subscribe("fal-ai/kling-video/v2.5-turbo/pro/text-to-video", {...})
   • Seedance 2.0 via FAL: fal_client.subscribe("bytedance/seedance-2.0/text-to-video", {generate_audio: False, ...})

3. Render overlay PNGs at target aspect via PIL:

   • Cold-open card: 3-line stacked text in Boska Black on transparent
   • End card: brand SVG logo rendered via rsvg-convert -w 2400 + autocropped → small annotations above/below in Space Grotesk + thin rule lines (Pinterest specimen-sheet style)

Phase 2 — Composite + music (autonomous)

1. Composite each variant via single ffmpeg filter_complex:

   • BG: trim/loop to duration, scale-to-fit-vertical + center-crop horizontal, apply per-variant dim filter (palette-aware — see references/composite-rules.md)
   • Cutouts: scale by WIDTH (not height — for 9:16) ~75% tall bottles / ~65% squat jars, overlay vertically centered (y=(H-h)/2), t-based enable='between(t,a,b)'
   • Cold-open card overlay at t≈1.5-3.0s
   • End card overlay at t≈7.8-10.5s
   • Output: h264 medium crf=20, yuv420p, +faststart, 30fps

2. Generate music bed via fal-ai/elevenlabs/music:

   • Brief: ambient instrumental, brand-category coded, 90-100 BPM, "builds subtly in last 2 seconds for end-card reveal"
   • music_length_ms: 12000 (generate slightly longer than video, trim — avoids 2nd-half decay tail)
   • output_format: "mp3_44100_192"

3. Process music — trim to duration + acompressor=threshold=-12dB:ratio=2:attack=20:release=200,loudnorm=I=-18:TP=-2:LRA=9 (EBU R128 -18 LUFS broadcast-ad standard)

4. Mux as SEPARATE ffmpeg pass per memory feedback_ffmpeg_lcut_endcard_recipe.md:

   ffmpeg -y -i master.mp4 -i music.mp3 \
     -c:v copy -c:a aac -b:a 192k -ar 44100 \
     -map 0:v:0 -map 1:a:0 -shortest \
     -movflags +faststart master-with-music.mp4
   

   Verify audio bit_rate ≈ 192000 via ffprobe (not 1000 — that's the default-mapping silent bug per feedback_ffmpeg_map_directive.md).

5. Self-QC via /watch:watch each variant:

   • Cutouts vertically centered (all SKUs at same visual mid-line)
   • BG dim appropriate (visible but not competing)
   • Cold-open card readable, dwells ~1.35s
   • End card readable, holds ~2.5s
   • Music instrumental, no peaks, audible
   • Duration ≈ target

Output

<project>/
├── brief.md
├── shot-list.yml
├── t2v-prompts-v3-kinetic.md
├── source/
│   ├── scraped-product-images/        # raw PDP PNGs
│   ├── scraped-brand/                  # logo SVG, palette
│   ├── scraped-brand-content/          # IG reels via gooseworks (reference only)
│   └── t2v-outputs/                    # N BG variants
├── assets/
│   ├── product-cutouts/                # birefnet-stripped PNGs
│   ├── fonts/                          # Boska + Space Grotesk TTFs
│   ├── text-overlays/                  # cold-open + end card PNGs
│   └── music/                          # raw + processed bed
└── finals/
    └── master-<aspect>-<concept>-<model>.mp4   # N variants

Decision Rules

1. Format wins over brand category for BG layer. Vignette = "kinetic but low-information." Real-time motion, NOT slow-mo. Even for clinical-luxury brands, the BG must be kinetic — slow-mo makes the ad feel sluggish.

2. Brand wins over format for typography + product cutout + end card. Use brand-appropriate typography (clinical-luxury → high-contrast modulated serif like Boska Black, NOT condensed industrial sans).

3. T2V skips grounding constraint. Text-to-video has no image input, so the "every image must be grounded" rule doesn't apply. Use T2V freely when stock fails. (I2V still requires a grounded keyframe — real photo or improvement of one.)

4. Scrape brand IG FIRST, before any Pexels search. Gooseworks scrapecreators IG endpoint costs ~$0.02 and returns brand-validated B-roll. Pexels for "luxury skincare macro" is a 5-pass dead-end — skip it after 2 weak passes.

5. Multi-variant comparison is the right spend for first-run. Fire 2 concepts × 3 models = 6 variants. ~$7-10 cost buys CERTAINTY about creative direction vs. guessing with 1 variant.

6. Cutout placement: vertical-center, NEVER bottom-anchor. Different SKU shapes (tall bottle vs squat jar) at bottom-anchor produce visually inconsistent placement. y=(H-h)/2 aligns all visual midpoints.

7. BG dim filter is palette-aware (per-variant). If BG has high-contrast elements competing with cutouts (chrome on cream), push saturation DOWN hard (saturation=0.50) — desaturating kills the metallic/vibrant character. For naturally-contrasty BG (dark ink on cream), lighter dim (saturation=0.85) is enough.

8. End card = annotated specimen-sheet, NEVER bare logo. Layout: small tracking text above (EST DATE) + horizontal rule + brand logo + horizontal rule + small annotations (ingredient name, novelty, key claim). Reads as brand specimen sheet vs slideshow exit. Per Pinterest research (Nolan Wells Studios + Daniel Soares film-credit references).

9. Music must be instrumental for Vignette format. Vignette has text overlays (cold-open + end card) — lyrics would compete. force_instrumental: true OR craft prompt with "instrumental only, no vocals, no lyrics".

10. EBU R128 loudness norm before mux. Raw music has variable loudness. acompressor + loudnorm I=-18 TP=-2 produces broadcast-ad-grade levels that sit cleanly under spoken VO (in case repurposed) and don't peak on consumer speakers.

11. ffmpeg mux: ALWAYS explicit -map 0:v:0 -map 1:a:0 + SEPARATE pass from composite. Default mapping silently produces 1 kbps garbage audio. Single-pass composite+mux complex filter silently fails (per feedback_ffmpeg_lcut_endcard_recipe.md).

12. Veo 3.1 quirks: aspect --aspect_ratio (underscore), aspect ONLY 16:9 or 9:16 (no 1:1), duration ONLY {4,6,8}, model veo-3-1-preview for premium quality, --quality ultra.

Quality Checks

Before declaring done, verify each variant:

• Duration within 9-12s
• Aspect ratio matches requested (typically 9:16 or 1:1)
• BG visible but does NOT compete with cutouts (per-variant dim correctly applied)
• All cutouts vertically centered (same visual mid-line across SKUs)
• Cutout alpha clean (no white halo)
• Cutout scale 50-75% frame width (NOT height for 9:16)
• Cold-open card readable, dwells 1.0-1.5s
• End card readable, holds 2.0-2.7s, includes annotated brand specimen-sheet style
• Music instrumental (no lyrics), normalized loudness (-18 LUFS ±3)
• Audio bit_rate ≈ 192kbps AAC (NOT 1kbps — check via ffprobe)
• No people, no hands, no products visible in BG
• No text/logos visible in BG
• Final mp4 plays cleanly in QuickTime + browser

Failure Modes

Failure	Cause	Recovery
Pexels searches return stock 3D abstract art repeatedly	"luxury / cosmetic / abstract" queries hit stock motion-graphics db	Skip Pexels after 2 weak passes. Go directly to gooseworks IG scrape + T2V
Cutout has white halo despite birefnet	BG color too similar to product edges	Re-run with finer matte settings or use NB2-edit ($0.08) to remove halo
Brand IG profile not on Instagram OR private	Some brands only have web presence	Skip Path E. Source BG via T2V directly.
T2V output has hard cuts when continuous wanted	"Shot N:" labels in prompt or multi_prompt arg	Strip "Shot N:" labels. Use "PHASE N" markers. Never set shot_type=customize
T2V output has slow-motion when kinetic wanted	Prompt phrase "slow motion" interpreted literally	Add "NOT slow motion" + explicit "real-time speed" + slow motion to negative_prompt
Veo 3.1 returns aspect error	Used 1:1 or --aspect-ratio (hyphen)	Use 9:16 or 16:9 with --aspect_ratio (underscore)
Music has lyrics	ElevenLabs ignored instrumental hint	Re-roll with explicit "instrumental only, no vocals, no lyrics" + brief regen. If 2nd attempt also fails, use a different prompt or stock
Music tapers in 2nd half	Known ElevenLabs Music behavior	Generate 20% longer than target, trim. OR per memory feedback_elevenlabs_music_decay.md — loop-and-flatten with acrossfade + acompressor + loudnorm
Final mp4 audio is 1 kbps	ffmpeg mux without explicit map	Add -map 0:v:0 -map 1:a:0 per memory feedback_ffmpeg_map_directive.md
Genesis (jar) sits visibly lower than Hero/Retinol (bottles)	Bottom-anchor overlay	Use y=(H-h)/2 (vertical center) for ALL cutouts
Chrome BG still competes after standard dim	Chrome highlights too prominent at default saturation	Push saturation lower (0.50 vs 0.85) — desaturating kills the metallic feel
Veo 8s output too short for 10.5s composite	Veo max duration is 8s	ffmpeg -stream_loop -1 -t 10.5 loops BG seamlessly; OR target 9s ad (matches V03 SLICK reference length)

Spend Reference (validated on Mother Science V-CARD, 2026-05-28)

Item	Count	Per-unit	Subtotal
birefnet-v2 product cutouts	3	$0.02	$0.06
Pexels Photos/Videos search	many	$0.00	$0.00
gooseworks IG scrape	1	$0.02	$0.02
Veo 3.1 ultra preview T2V (8s)	2	$2.00	$4.00
Kling v2.5 Turbo Pro T2V (10s)	2	$0.70	$1.40
Seedance 2.0 T2V (8s 720p)	2	$0.85	$1.70
ElevenLabs music (12s)	1	$0.40	$0.40
ffmpeg composite + mux	6	$0.00	$0.00
Total (6 variants comparison)			~$7.58

For single-variant runs: roughly $1.50-2.50 depending on model.

Sub-Archetypes

• V-CARD (validated default) — Cold-open text card (1.5s) → product carousel (3 SKUs × 1.5s) → annotated end card (2.5s). 10.5s total. Use when brand has hook copy + multi-SKU lineup.
• V-PURE — Pure cutout carousel, no cold-open, no end card text (logo only). Most minimal. Use when product visuals alone carry the message.
• V-TILE — Persistent social-tile layout with headline + quote + rating + logo over BG. Use when brand has strong review/quote content. Heavier production (custom layout, possibly animated bloom endcard).

This molecule defaults to V-CARD. V-PURE = remove cold-open beat + simplify endcard. V-TILE = significant restructure (see references/format-dna.md for the V02 ClimbOn reference pattern).

Tests

See tests/:

• smoke-test.md — structural pass
• sample-input.md — Mother Science worked example
• expected-output.md — what successful output looks like
• human-test.md — manual acceptance criteria
• verifier.md — ffmpeg + ffprobe assertions

Related

• Reference project: clients/mother-science/video-01-clinical-claim-vignette/ — full worked example with HOW_TO_MAKE_THIS_VIDEO.md + LEARNINGS.md (20 load-bearing learnings)

• Related molecules:

  • skills/molecules/ugc-ad/* — for UGC creator ads (different format)
  • skills/molecules/product-video/* — for product-anchored ads (different format)
  • skills/molecules/explainer-video/* — for VO-driven explainers (different format)

• Memory rules to read before running:

  • feedback_ffmpeg_map_directive.md — -map 0:v:0 -map 1:a:0 always
  • feedback_ffmpeg_lcut_endcard_recipe.md — separate composite + mux passes
  • feedback_elevenlabs_music_decay.md — generate longer + trim
  • feedback_image_grounding_rules.md — applies to I2V keyframes only, not T2V
  • reference_higgsfield_veo_cli_flags.md — Veo aspect/duration constraints
  • reference_kling_v3_pro_i2v.md — Kling endpoint reference (note: this molecule uses Kling v2.5 Turbo Pro T2V, not v3 i2v)
  • feedback_endcard_split_frame_layout.md — alternative endcard pattern
  • feedback_endcard_reference_pattern.md — endcard via reference + NB2 + HTML

Open Questions (not yet Decision Rules)

These need N=2+ project validations before promoting:

• Optimal BG concept count per run — we did 2; would 1 (cheaper) or 3+ (more variety) be a better default?
• Per-aspect cutout scale defaults — we validated 75%/65% width for 9:16. Are 1:1 defaults the same or different?
• Music decay frequency — ElevenLabs decay didn't manifest in our run; may be brief-specific. Track on next 3 brands.
• gooseworks IG scrape reliability — worked perfectly for Mother Science (US prestige brand, active IG). Coverage for India market / niche brands TBD.
• T2V brand-IP risk — we generated chrome molecular spheres + black ink in cream that are very SIMILAR to what Mother Science already uses. Is "inspired by brand language" an issue for brand IP? Surface to legal review for first 2-3 brand runs.