-
-
Your own hair, back at week 4. eCommerce VTO: wigs, headwear and brows on one photo, stock ranked against her kept hair, a live try-on log.
-
Home page: her own hair back on her week-4 photo, with the photo she kept as the reference. Anjali is a demo person generated with YouCam.
-
A shop's own wig photo tried on her face: Hairstyle Try-On with the product photo as reference, run live (2 units; 2 of 2 runs worked).
-
Regrowth, read against the hair she kept: Hair Type and Hair Length on new growth (a generated example, not a prediction).
-
The fitter desk ranks a shop's own stock against the client's hair, showing each score's parts and a hold note per wig. 0 YouCam units.
-
Ask the fitter desk: a rules-based planner walks Hairloom's MCP tools to a pull list. No model calls, no YouCam units.
-
The Try step at phone width: the mirror and the named wig rail on one screen.
-
Colour measured on the device, with its share of grey: Margaret reads about 27% grey and gets a blend code to check against swatches.
-
Alopecia: a Hairstyle Try-On covers the bare patches; a Hair Volume try-on leaves them showing (generated photos; Nia is a demo person).
-
Not only chemo for women: Ravi's own hair back, Arjun (a men's wig) and a flat cap, on his week-4 photo (Ravi is a demo person).
-
How it's built (Next.js 15, 5 API routes): photos stay on the device, the server whitelists YouCam tasks, Skin Analysis sets the cap note.
Donna chose her wig before chemotherapy, at the first fitting, and still needed a second one: when she collected it, it didn't look right to her (Beatson Cancer Charity). Hairloom shows the wig on your own face after hair loss before anyone buys: your own hair back on your week-4 photo, the shop's stock ranked against the hair you kept, and one sealed brief for your fitter.
Which topic: eCommerce VTO. A wig, headwear and brows are tried together on the same week-4 photo, and the shop's own stock is ranked against the hair the client kept. Skin AI sets the fitter brief's cap-lining note.
What Hairloom does that a wig catalogue can't
- Your own hair, back on your week-4 photo. Hairstyle Try-On, with the photo you kept before treatment as the reference.
- A sealed brief for your fitter. One encrypted link carries the shortlist on your face, your shade code, cap size and skin trend; the fitter desk adds real listings near each wig.
- The shop's own stock, ranked against the hair you kept, with a hold note per wig ("Please hold KC-ASHA-2 (Asha) in shade 4, cap Average, for Anjali (demo)"), on the fitter desk or over MCP. A shop's own product photo can be tried on too.
- Regrowth, re-read against the hair you kept (on a generated example, not a prediction). Kept: type 1 to 2A, above chest. Now: 3B to 3C, short.
- Not only chemotherapy: alopecia areata, pattern thinning and men. 11 YouCam APIs, and every task we kept in a public ledger, failures included.
Inspiration
About 65% of people on chemotherapy lose their hair, and fear of it can lead some to refuse treatment (JAAD). In a survey of 1,478 women after chemotherapy for breast cancer, 84% wore a wig at first and 47% still did a year after treatment (PLoS One, 2019). The American Cancer Society suggests photographing your hairstyle before treatment to match a wig (ACS); Hairloom keeps those photos and reads them. The real wigs we ranked cost £156 to £520 or $155 to $639, and a worn wig can rarely be returned.
How it works
- Your kept-hair record. Three photos while your hair is still yours, read by YouCam (curl, frizz, length, face shape and brows), with head size for the cap and a colour measured on your device, including its share of grey.
- Try wigs before and after. At week four, Hairstyle Try-On puts your own hair back from the kept photo. Every wig in the boutique is ranked against the hair you kept, with the score's parts shown (length and texture steps, colour ΔE00), and any of them can be tried on your face (2 units each live; all are ready for the demo people). Hair Color Try-On recolours one to your shade.
- Brows, headwear and skin. Makeup Try-On rebuilds brows; headwear uses the Image Generator with an instruction to keep your face. Skin Analysis check-ins each cycle set a cap-lining note in the fitter brief when they read more sensitive (cosmetic scores, not a diagnosis).
- One sealed link to your fitter, encrypted in the browser with the key only in the link.
- Read the new growth. Hair Type and Hair Length on one shared upload (4 units), against the kept record.
The four demo people are Anjali (breast cancer), Nia (alopecia areata), Margaret (grey-blonde, a taxane regimen) and Ravi (a man with lymphoma). Each opens free with a ready shortlist and a fitter brief at /fitter#demo=nia and so on. The fitter desk also ranks 13 real wigs from two retailers' public product pages (metadata only, no partnership), and "Ask the fitter desk" walks the MCP tools to a pull list with no model and no units.
The 11 YouCam APIs, and what each does here
Run counts are from the evidence page, which lists every task we kept: 270 runs, 255 succeeded.
| YouCam API | Job in Hairloom | Runs |
|---|---|---|
| AI Hairstyle Virtual Try-On | Every wig on your week-4 photo; your own hair back; a shop's own photo | 98/98 (96 wig and own-hair + 2 shop-photo; 18/18 on real photos) |
| AI Hair Color Virtual Try-On | The wig in your measured shade code | 19/19 |
| AI Hair Type Detection | Curl pattern, for the record and the regrowth re-read | 6/6 |
| AI Hair Frizziness Detection | Frizz, from the same three photos | 5/6 |
| AI Hair Length Detection | Length, which ranks wigs | 6/6 generated, 2/5 real |
| AI Face Attribute & Ratio Analyzer | Face shape and brow shape and colour | 0/5 real at default strictness, 5/5 at "flexible" |
| AI Makeup Virtual Try-On | Rebuilt brows | 9/9 |
| AI Image Generator (image to image) | Headwear, asked to keep the face; the demo's later photos | 53/53 |
| AI Hair Volume Virtual Try-On | Fuller hair on a photo of thinning hair | 9/9 |
| AI Hair Extension Virtual Try-On | Clip-in length | 8/8 |
| AI Skin Analysis | Cosmetic skin check-ins that set the brief's cap-lining note | 3/3 |
A shop's own product photo as the try-on reference (Hairstyle Try-On, its own whitelisted kind) ran 2 of 2 through the deployed studio. Hat and Scarf Try-On were tested and left out: Hat restyled the whole person and added hair under a beanie on a bald head, and Scarf draws a neck scarf, not a head scarf, so headwear uses the Image Generator.
How it's built (Next.js 15 on Vercel, code as judged at tag v1.4-submission): photos and the record stay in IndexedDB on the device. The browser names a task kind and the server builds the body from a whitelist, checks a daily unit counter, reserves units before any await and returns an HMAC-signed handle, never the task id. Each task is deleted at YouCam after its result, with a capped sweep and a daily cron for anything left. Colour and grey are measured on the device, fitter links are sealed with AES-GCM, and unit costs are compared kind by kind with YouCam's feature-cost list (13 of 13 match, checked Oct 30 at no cost); a test fails the build if they drift. The lab scripts that made every demo asset are in lab/.
What we measured, including what failed
- Bald heads: YouCam's docs list
error_hair_too_short, yet on the four demo people's bald photos 64 of 64 wig try-ons and 4 of 4 own-hair restorations worked, and so did 8 of 8 try-ons on four real shaved heads (counts published; photos and try-ons kept private). - Light shades: Margaret's three blonde recolours landed a median ΔE00 17.1 from the asked shade at colour intensity 70, 11.0 at 100, and 5.0 when the colour sent is also lifted by the measured shortfall. That lift was fitted on these three shades (in-sample), so light shades still say to check a swatch. CIEDE2000 is tested against the Sharma, Wu & Dalal reference pairs.
- Headwear and the face: across 18 renders, face keypoints moved a median 0.049 eye-widths (at most 0.091); not a face-recognition match.
- Bare patches: a wig try-on covered them; a volume try-on left them showing.
- Real portraits: face reading 0 of 5 at the default angle strictness, 5 of 5 at "flexible"; hair length 2 of 5; on one bearded man, wig try-ons removed the beard. Hair Density rejected all 6 photos for their angle, so it isn't in the studio.
Not proven yet
- No real person with hair loss has used Hairloom from start to finish, and no shop has used the fitter desk (no partnership with the retailers listed).
- The bare patches were generated and drawn pink, unlike real alopecia areata. Camera Kit hasn't been tested on a real phone.
- Grey blend codes and the ranking haven't been checked against swatches or against what fitters and clients choose.
What's next
An eight-week pilot with one oncology wig bank or boutique in Pune or Mumbai; patients never pay (who does).
Why this fits the event
eCommerce VTO where it matters most, since a worn wig can rarely be returned: stock ranked against the client's own hair and tried on her face, a shop's own photo tried on, real listings and hold notes. Skin AI check-ins set the brief's cap-lining note and the MCP plan. The build above adds an MCP server and 55 tests.
Built during the submission period (Oct 2026). The code is MIT. The demo people and wigs are generated with YouCam and are not real customers or products. Hairloom helps with appearance, not medical care, and never diagnoses.
Built With
- google-cloud
- mediapipe
- model-context-protocol
- next.js
- react
- typescript
- vercel
- web-crypto
- youcam-api
- zod
