Evidence
246 YouCam tasks are in the ledger below, and 235 succeeded: 177 made the demo people and boutique stock, 41 ran on real photographs, and 28 were early research. Times are wall-clock from creating the task to having the result, polled every few seconds.
The studio calls 11 YouCam APIs, and each one has at least one successful run here. Every row is counted from an output file we kept (2026-10-21, scripts/export-runs.mjs), so tasks whose output we didn’t keep, such as early exploration and one-unit end-to-end checks, are not in it. Face and hair length are split because they behave differently on generated and real photos.
The demo people are generated, so we ran the same YouCam calls on nine real portraits from Unsplash. Four are shaved heads. These are Unsplash-licensed stock photos of models, credited to each photographer on its card, and the people are not patients. We publish the results, not the pictures.
Long straight brown hair, sunset backlight
Tested on a photo by Štefan Štefančík on Unsplash; 2 try-on images kept privately.
Long curly hair
Tested on a photo by Jude Infantini on Unsplash; 2 try-on images kept privately.
Long grey hair, smiling
Tested on a photo by Soroush Sedghi on Unsplash; 2 try-on images kept privately.
Short dark hair
Tested on a photo by Jurica Koletić on Unsplash; 2 try-on images kept privately.
Curly bob, warm light
Tested on a photo by Filipp Romanovski on Unsplash; 2 try-on images kept privately.
Shaved head, beard
Tested on a photo by Zoran Borojevic on Unsplash; 4 try-on images kept privately.
Shaved head, smiling
Tested on a photo by Ben Masora on Unsplash; 4 try-on images kept privately.
Shaved head, black and white
Tested on a photo by Lisa Marie Theck on Unsplash; 4 try-on images kept privately.
Shaved head, full beard, outdoors
Tested on a photo by Tekeyah Whit on Unsplash; 4 try-on images kept privately.
On the real photographs above, YouCam’s label fit the hair on 2 of the 3 photos where it returned one, and Hairloom’s measured shade gives a code a fitter can order. Both are judged by eye, because we have no reference swatch yet, and the measured shade missed once, on backlit brown hair.
For each generated demo person, the colour label YouCam’s Face Attribute Analyzer returned for the kept photo, and the shade Hairloom measured from the hair pixels of the same photo. The label was “Blonde” for all 3 generated people that got one, whether their hair was dark brown, blonde or jet black. That looks like an artefact of generated images, so where it agrees with the measurement below, that is chance, and we report it rather than rely on it.
They disagree, so the passport uses the measured shade.
Nothing to compare.
They agree, by chance.
They disagree, so the passport uses the measured shade.
Each demo wig was generated with a shade code in its description. Measured from the product photo, 5 of 16 come back as that code; the rest read darker and less red, by 4.0 points of lightness on average, because shadowed strands pull the median down. So the studio and the fitter desk show the listed code as the stock shade, and use the measurement only to rank wigs against each other. A client’s measured shade carries the same bias, which is why the passport and the brief ask the fitter to confirm against real swatches.
| Wig | Listed | Measured | ΔE00 |
|---|---|---|---|
| Asha | 2 | 1B (miss) | 5.2 |
| Meera | 1B | 1 (miss) | 7.6 |
| Nora | 6 | 4 (miss) | 7.6 |
| Ivy | 30 | 6 (miss) | 17.5 |
| Zuri | 1B | 1 (miss) | 8.8 |
| Amara | 1 | 1 | 5.1 |
| June | 38 | 10 (miss) | 8.6 |
| Silver | 56 | 51 (miss) | 6.0 |
| Rosa | 33 | 6 (miss) | 9.4 |
| Lena | 44 | 44 | 4.5 |
| Tara | 4 | 4 | 3.5 |
| Wren | 38 | 38 | 4.5 |
| Kai | 4 | 1B (miss) | 7.9 |
| Dev | 44 | 44 | 8.9 |
| Arjun | 1B | 1 (miss) | 11.1 |
| Sam | 4 | 1B (miss) | 9.9 |
Each note is something that cost us a run, what Hairloom does about it, and the one change we’d ask the Perfect Corp team for.
Front, then the face turned towards the right of the picture, then towards the left. Any other order, or different sizes, returns error_face_angle_invalid or error_mismatch_image_size. The studio crops all three to the front photo's size and orders the side photos itself from face landmarks, so a swapped upload still reads.
We’d ask: Say which photo failed, or accept any order and size.
Face Attribute Analyzer and Facial Color Tones Analyzer return one of a few named swatches. That suits a beauty quiz, but a wig fitter needs a shade code, so Hairloom measures colour from the hair pixels and keeps YouCam's label beside it. Where they disagree is listed below.
We’d ask: Return a measured hair colour with a confidence, as eyebrow_color already is.
The docs list error_hair_too_short, so we expected failures after hair loss. All 64 try-ons on bald demo photos succeeded, and so did all 8 on real photos of shaved heads. Those are smooth scalps in good light, not patchy loss, so patchy and stubbled heads are still untested.
We’d ask: Document that bald and shaved sources work.
At the default strictness, Face Attribute Analyzer rejected all five of our real, everyday portraits (error_face_not_forward_facing, error_face_angle_left_tilt, error_face_angle_downward). At face_angle_strictness_level “flexible” all five read, and face_quality still flags a turned face, which the passport passes on. Hair length has no such setting and failed on three slightly turned portraits, which is why the studio leads with YouCam's guided camera.
We’d ask: Name face_angle_strictness_level in the error, and give hair length the same setting.
On one real photo of a man with a full beard, both wig try-ons removed the beard, and on a black-and-white portrait the try-ons came back in colour. Hairloom's mirror lets you drag back to the original to spot this, and a fitter should always see the real person.
We’d ask: A hair-only mode that leaves beard, skin tone and background alone.
Hat and scarf try-on generate a new outfit and background, and on a bald head the beanie came back with long hair under it. For headwear Hairloom uses image-to-image with an instruction to keep the face, clothes and background, at 1 unit.
We’d ask: A keep-identity option on hat try-on, and a head-scarf style.
The same front photo read the left brow tail as “Average” on one run and “Sparse” on another. The passport keeps the reading with the photo it came from rather than presenting one run as fact.
We’d ask: A confidence value on shape and thickness readings.
Front-facing and generated head-down photos were rejected with error_face_angle_invalid. It needs YouCam's camera guide in hairdensity mode, so it is left out of the upload flow. For hair loss this is the reading we missed most, because the passport can say how long hair is but not how much of the scalp it covers.
We’d ask: A scalp-coverage readout that works on an ordinary front photo, for alopecia and thinning.
Every hair-volume and hair-extension template in YouCam's template list is in the “Female” category, so Hairloom doesn't offer them on men's photos. On a generated photo of thinning hair with a widening parting, all four volume templates filled out the sides and added length, but none covered the parting or a thin crown, which is what a topper is for. The three length templates changed length only.
We’d ask: Men's templates, and a crown or parting topper template.
The face must fill most of the frame. The journal finds the face on the device and crops to it before sending, which turns a normal selfie into an accepted one.
We’d ask: Accept a normal selfie and crop to the face on YouCam's side.
Tested on Oct 21: after task/delete the result URL returns 404, and starting a new task with the same input file fails with a 403. Hairloom deletes each task as soon as the browser has the result. If the tab closes first, the page sends the delete as it closes, and a daily sweep deletes anything left over.
We’d ask: Delete on completion as an option, so no task depends on the browser to clean up.
Wig shade codes are a loose convention that varies by brand; our hex values approximate typical swatches.
The real-photo tests use Unsplash photos, under the Unsplash License and credited to each photographer. They are not covered by the code’s MIT licence. MediaPipe models are Apache 2.0. See the NOTICE file in the repository.