# Wonder: the change that improved v50

This is a truthful process handoff, not a claim that the model learned permanently or that the ads have proved performance. The operator preferred v50 over the preceding versions. No spend test has established a winner.

## What changed

1. **Inspect the entire competitor ad.** Read headline, type, line breaks, color contrast, photo, product staging, offer, review count, stars, guarantee, supporting line and margins. Obvi O06/H04/O28 informed the current work. Earlier attempts copied the main headline/product idea but missed much of the supporting argument.
2. **Specify the finished ad, not an empty scene plate.** Earlier versions generated photography then added deterministic Remotion typography. v49 already generated whole photographic scenes, but the operator still thought the bottle looked pasted in. v50 asked the image model to generate photography, bottle, lighting, label perspective, headline, price badge and supporting graphics together in a single call. No separate product insert or later compositing.
3. **Use the real package as identity reference only.** Explicitly ask the model to re-render it from the scene's camera angle, with shared light, contact shadow, realistic scale, fingers/occlusion and label foreshortening. Do not extract the bottle pixels or place a flat packshot over a scene. Check the actual result: prompt wording alone does not prove integration.
4. **Give typography exact instructions.** Huge, ultra-heavy, wide geometric sans-serif; short intentional line breaks; a restrained two-color headline; cheerful saturated palette; explicit exact text. The successful v50 directions were pink/plum, coral/wine and cobalt/yellow. This is visual prompt direction, not a verified font-file identity.
5. **Give each sales element a job.** Growth/regrowth is the desire; the hair photograph shows that desire; reviews supply trust; the $14 first subscription bottle lowers the purchase barrier. Do not let ingredients substitute for reviews. v50 still missed the review stack; v51 corrects it. A big bottom bar is optional, not a requirement copied from Obvi.
6. **Design for a specific placement.** These are 4:5 feed concepts. Reserve breathing room around key text and product labels, then inspect the actual generated pixels. A house margin or illustrative CTA overlay is not Ads Manager validation. Do not present feed assets as Stories/Reels-ready.
7. **Preserve provenance and test phone legibility.** Save the full prompt, official reference, original tool output path and hash. Keep final originals byte-identical. Only resize/compress review previews. View the actual compressed result at phone width and check every word, numeral, star/count, hand, bottle and edge. Regenerate the whole image if it fails; do not quietly repair it with pasted elements.

## Copyable prompt structure

> Generate ONE complete finished 4:5 advertisement, including scene, people, bottle, label perspective, lighting, typography, review panel and offer. Use the official bottle image only for identity and re-render it naturally in this scene. No product cutout, flat pasted packshot, separate scene plate or later overlay. Use [specific palette], ultra-heavy wide geometric sans-serif, exact headline [TEXT / LINE BREAKS], and [specific scene]. Include [verified, correctly attributed review stack] and [exact offer + qualifier]. Place all important text, faces and bottle labels within [explicit composition regions]. Continue the background below the protected content area, without a giant footer. Spell text exactly. No invented reviews, result timelines or guarantees.

For comparison mockups: specify the same fictional adult in both views, matched camera, pose, lighting, exposure and clothing, with the intended hair-density change only. Document them as generated placeholders on the surrounding review board; they are not real customer evidence.

## Evidence and remaining limits

The depth and occlusion in the complete image generations are model-generated, not a later hand-cut mask. In v51, percentages alone did not reliably move content upward. Replacing the vertical order did: headline, then reviews and price, then comparison photos, then quiet background. Three complete replacements corrected low supporting content, and another corrected a duplicated attribution word. Inspect the output rather than assuming the prompt was followed.

- Accepted visual direction: https://vps-327b7201.vps.ovh.us/WB/ADS/wonder-v50-20260922/
- Exact v50 prompts: `research/meta-statics/2026-09-22-v50/brief/generation-prompts.json`.
- Original identity/hashes: `research/meta-statics/2026-09-22-v50/brief/lineage.json`.
- v51 prompts with review stack: `research/meta-statics/2026-09-22-v51/brief/generation-prompts.json`.
- The company Trustpilot score is brand-level, not Wonder efficacy proof. Do not invent a hair-growth customer quote or turn a generated comparison into a customer result.
- This method improved operator preference. It does not establish conversion performance, package-artwork exactness, claim substantiation or Meta approval.