Beauty Studio - Glow Up AI

Beauty Studio is an AI beauty assistant that analyzes your features, recommends the best hair, colors, makeup and style, and lets you preview your new look.

Comment

Inspiration

We wanted one photo to be enough: a report that reads like a stylist actually looked at you, and a way to try a new haircut, hair color or makeup look on your own face before committing to it.

What it does

Beauty Studio turns one portrait into a personal beauty report and photoreal try-ons.

  • 92 analyses across hair, color, makeup, outfits, skin, nails, face and style
  • Written reports such as face shape, undertone and seasonal palette
  • Try-ons for haircuts, hair colors, makeup looks, outfits and manicures
  • Reports in 9 languages, in the tone you choose, and you can correct a wrong reading
  • No account needed to start or to buy; Apple, Google or email can be linked later

How we built it

  • Expo and React Native with TypeScript
  • Supabase for auth, database, storage and edge functions
  • fal.ai models called only from the server: Gemini 2.5 Flash for reports; Nano Banana 2, FLUX.2 Pro and Nano Banana Pro for try-ons, with an automatic quality check on every image
  • RevenueCat for three subscriptions (weekly, monthly, yearly) and four credit packs. Purchases are confirmed on our server through RevenueCat webhooks, and credits are charged and refunded server-side.
  • The Supabase user id is the RevenueCat app user id, so guests can buy right away and keep their purchases when they create an account

Challenges we ran into

  • Keeping the person the same person. A try-on is only useful if you still look like yourself. Every try-on prompt locks identity (and hands for nail try-ons), each preset is limited to one scope such as hair color or outfit, and a second vision model rejects renders that change more than that scope allows. Rejected images are re-rendered on the next model with a new seed.
  • Visual consistency across a session. In the Full Visual Session each step (hair, color palette, makeup, outfit, glow-up) is rendered on top of the previous step's result, so the look builds up instead of starting over. Outputs identical to the source photo are caught by hash and re-rendered.
  • Shipping 92 analyses without chaos. Every analysis is a declarative entry (type, category, credit cost, cover image, promised outcomes) with a blueprint of typed report blocks. Generators copy them to the server, contract tests check every entry, and a QA mode renders all 92 screens with mock data without spending on AI. Each one ships in 9 languages.
  • Reports that are true to the photo. Every sentence has to pass four tests: it would not fit anyone else, it names a mechanism, it measures in fingers, centimeters or tone levels, and it never repeats. Colors are read from the photo's own light, scores and medical claims are banned, and users can correct a reading to have the report rewritten.
  • Payments that arrive later. The store charges instantly but the RevenueCat webhook lands a moment later, so the app keeps a silent watch until the credits appear, and every grant is idempotent.

Accomplishments that we're proud of

  • 92 analyses from a single photo: including 47 photoreal try-ons and 3 multi-step sessions, all in 9 languages
  • A design that feels like a beauty magazine: dark ink and champagne gold, editorial serif headlines, chapter-style reports with face diagrams and palette swatches, and a unique cover for every analysis
  • AI art direction that matches the app: every analysis cover comes from one shared style brief, so the whole catalog reads as one collection
  • Try-ons that look real and still look like you: identity locked in every prompt, one change per try-on, and an AI quality gate that re-renders anything that drifts
  • Sessions that stay consistent: multi-step looks build on each previous render instead of starting from scratch
  • Reports that read like they were written for you: grounded in your photo's light and proportions, in the tone you pick
  • A monetization flow on RevenueCat: guests subscribe without signing up, grants are idempotent, and failed jobs are refunded automatically

What we learned

  • Consistency comes from constraints, not longer prompts: lock identity, allow one change per try-on, verify with a second model, and chain steps on the previous result.
  • Image generation is a queue, not a request: processing leases, attempt caps, a seed per attempt, hourly cleanup and automatic refunds keep long renders reliable.
  • Honesty has to be measured: rerun the same photo to check that a reading is stable, and compare different photos to catch generic text.
  • One source of truth for data: the analysis registry generates report blueprints, server copies and credit maps, and CI fails when anything drifts. Reports are typed JSON with a schema version, so old reports upgrade instead of breaking.
  • Private by default: photos live in per-user private storage behind signed URLs, and users can really delete their photos, reports and account.
  • Money data belongs on the server: RevenueCat owns the payment, our database owns the balance, and every grant has an idempotency key.

What's next for Beauty Studio - Glow Up AI

  • AI Photo Studio: beauty-aware editing and natural retouching for your own photos. Apply the looks you loved to any picture and stay recognizably you.
  • Your personal beauty guide: a coach that turns every report into a living plan, with daily routines, occasion looks and reminders that adapt to what you save.
  • Smart recommendations: makeup shades, hair colors, jewelry metals and outfits matched to your palette, undertone and face shape.
  • Glow-up tracking: follow your progress across analyses with side-by-side comparisons.

Built With

Share this project:

Updates

Submission history