Labelle

Inspiration

Every phone gallery is a landfill of identical photos. We missed the way film made each frame cost something — you posed, you waited, the print came out, and it was yours. Labelle rebuilds that ritual: a little print shop in your pocket where every AI edit is a PRINT with a №, developed on a press, hung to dry — and where the person in the photo is always still you.

What it does

Booth — a film camera with 22 in-house film stocks: parametric color science, live animated plate grain at film cadence, spectral halation (red halos wide, blue tight), 1/2/4-up collage layouts. Studio — YouCam-powered tool rooms: İYİLEŞTİR (AI Photo Enhance), RENKLENDİR (Colorize), KALDIR (brush-mask Object Removal with one-tap Vision subject pick), FON (Background Replace), and SET (AI Studio) — plus GERÇEK IŞIK, our fully on-device relight engine (Depth Anything V2 + person matting → real directional light with contact shadows), with a postage-stamp animation that shows where the light will come from. Beauty — the full YouCam face catalog on one icon rail: skin, reshape, eyes, smile, hair, body, glow — one APPLY compiles the fewest billable calls, and a Skin Report tab runs the 8-concern analysis. Closet — shoot a garment IN A STORE, the cutout goes through Salient Object Detection, and cloth-v4 puts the real piece on your photo; AYNA is a live mirror with a front/back camera flip where each try is one still, seconds later, on you. Darkroom — the roll. Local-only editing: film, exposure dials, crop, collage. Every AI print passes an identity check (Vision face distance, framing-calibrated) before it lands — if it isn't you, it doesn't print, and it doesn't charge. Six languages (EN/TR/DE/FR/ES/PT-BR) driving copy, dictation locale and casing; every refusal in the app is visible and priced in PRINTS, never silent.

How we built it

SwiftUI + Metal on iOS 26, orchestrated as a fleet of AI coding agents over ten sprint rounds with a single non-negotiable gate: verify.sh — emoji lint, identity-threshold calibration against tracked fixtures, then the full suite. The submission tree closes at 1,444 tests across 249 suites, all green. The YouCam integration is a typed client over the S2S task lifecycle (auth → upload → task → poll → download) with per-action poll ceilings, a live price book pulled from feature-cost, a unit ledger that adopts the vendor's real balance at launch, and a credit gate that authorizes a whole billing chain before the button press.

Challenges we ran into

The honest list is long — the three best are in the dedicated form answers below: the /pre-process coordinate-space surprise, the billed-no-op discovery in appliedFeatures, and the silent-refusal hole our own paywall dug. One more worth naming: making grain feel like film — we ended with tiling "grain plates" sampled per tonal band (the way pro film emulators use real emulsion scans), animated at 16fps in the viewfinder and frozen at bake, because a print is paper.

Accomplishments that we're proud of

An identity gate on every AI print — the app refuses (free of charge) any result that stops looking like you, with thresholds calibrated, not guessed (same-person re-framing noise measured at three framings against tracked fixtures). A fully on-device relight engine that pairs with YouCam's cloud tools instead of competing with them — depth, normals, contact shadows, 47.5 MiB, Apache-2.0. A billing model that cannot lie: chains authorize up front, the ledger commits once, a failed print never burns a credit, and "free" is silent by law.

What we learned

That a metered API deserves the same engineering respect as a database: read the real price book at runtime, measure per-call billing with a stopwatch, and design every refusal to be visible. And that vendor docs are a map, not the territory — two of our biggest fixes came from disbelieving a doc and running one cheap live call.

What's next for Labelle

App Store release (keys move behind a proxy first), the trial + print-pack economy that is already wired, and hair/nail live AR when our vendor budget graduates from hackathon grant to production plan.

Built With

• avfoundation
• core-image
• coreml
• depth-anything-v2
• metal
• perfectcorp
• storekit
• swift
• swiftdata
• swiftui
• vision
• xcodegen
• youcam-api