-
-
The bought blouse beside two replacements on the same synthetic person. Genuine recorded YouCam Clothes V4 results, with prices & labels
-
The sage preview's measured colour differs from its product photo. Swatches expose the difference; checks are warnings, not proof of fit.
-
Inspect detected image changes: blue inside the fixed comparison band, orange outside it. The overlay is a pixel comparison.
-
Beauty and fashion in one order: Clothes V4, Makeup VTO and Necklace VTO applied in sequence on another synthetic demo person.
-
An even exchange has no match. A proposed $2 budget change leaves other limits intact and takes effect only after confirmation.
-
Review $40 credit, a $42 replacement and a $2 top-up before creating a request. This fictional quote reserves no stock and takes no payment.
-
The request continues into staff review: a question and shopper reply remain on the shopper's page. Fictional demo dialogue.
-
Registered study results, including failures and limits. These are measured pass counts and blind AI labels.
Try the blouse exchange · Watch the 2:50 demo · Judge guide · YouCam results and evidence
Try this first: choose I don't like how it looks, confirm a quieter print, long sleeves and at most $10 extra, then Preview on a person using recorded Model B results. Compare the bought blouse with the replacements. The sage top has a Flagged: colour warning: inspect it, choose whether to proceed, then review the exchange price. No login is needed.
For judges: select Judge access and enter the code in the private submission attachment for capped live YouCam try-ons, own-photo uploads, staff review and AI planning. Unlocking itself makes no paid call. Source code and local setup are public. The judge code is only for the hosted live features and staff tools.
Inspiration
The blouse fits. The print is too busy. You would exchange it for something quieter with long sleeves, but another product photo leaves you guessing whether the replacement will be any better.
Rechoice brings virtual try-on to that moment after purchase: which replacement addresses the reason for this return, within the shopper's limits and the store's policy? The original purchase becomes the comparison baseline. Shoppers can inspect the change before requesting an exchange; staff receive the selected item, price and shopper's reasons together.
What it does
- Explain the return. Open the bought blouse and confirm long sleeves, a quieter print and at most $10 extra. Catalog facts, stock, price and return policy determine eligible replacements.
- Compare on the same photo. YouCam Clothes V4 renders each item on the same photo; Rechoice displays the results side by side. Where inputs are available, local colour, print-detail and image-stability checks add visible warnings. Colour swatches and Show what changed help inspect a concern. A flagged preview needs explicit acknowledgement before a quote; the app rechecks that acknowledgement when saving the request.
- Explore a trade-off. If an even exchange has no match, review a proposal to allow $2 extra for a quieter long-sleeved top. Other confirmed limits stay intact. Reviewing changes nothing until the shopper confirms.
- Request an exchange. Review the price, send the request, answer a staff question and continue through staff review. A normal return remains available throughout.
For the eCommerce VTO theme, See it with the rest of my order connects beauty and fashion: Clothes V4 → Makeup VTO → Necklace VTO shows the replacement with the order's lipstick and necklace. Each output becomes the next stage's input. The final combined image is labelled unchecked.
A retailer API creates order-specific return links, while staff insights show searches the catalog could not satisfy. This gives stores a concrete way to explore suitable exchanges and unmet demand.
How we built it
Django serves the shopper and staff flows, with PostgreSQL, Redis and Celery. Workers upload through YouCam's File API, create tasks, poll for completion and store results privately. A persistent unit ledger reserves allowance before generation and keeps uncertain outcomes accounted for.
The core APIs have checkable results in our original pilots:
| YouCam API | Role in Rechoice | Completed pilot tasks |
|---|---|---|
| Clothes V4 | Bought-item comparison and replacement clothing | 10/10 |
| Makeup VTO | Lipstick and foundation; lipstick in the combined look | 10/10 |
| Necklace VTO | Necklace alternatives and combined look | 4/4 |
These are task-completion counts, not visual-quality scores or counts of combined looks; one necklace output was rejected. Nail VTO, Earrings VTO, AI Hair Color and Shoes V3 also have working integrations and published results.
The top checks use CIEDE2000 colour distance, relative edge detail and changed pixels outside a fixed band, with no extra model call. An optional Claude copilot uses catalog and eligibility tools to propose plans; confirmed shopper controls remain authoritative.
Challenges we ran into
Successful generation did not always produce a useful comparison. A sage top rendered lighter than its product photo. A smaller earring looked as long as the bought pair; we added a relative-size check and a labelled correction while retaining the original. Bag V2 changed the outfit, pose and background, so we disabled bag previews.
The checks also make mistakes: pockets can resemble print detail, lighting affects colour measurements, and a fixed band is not garment segmentation. Warnings cannot guarantee fit or exact appearance.
Accomplishments that we're proud of
Two studies registered before rendering produced 46 live YouCam results with check thresholds fixed in advance. 15 of 26 plain-top alternatives passed all three checks; the other 11 were flagged for review. All six repeats of three colour-flagged cases stayed flagged. These are check outcomes, not accuracy. The evaluation includes images, measurements and blind AI labels, including disagreements.
The final release passed 441 tests on each of PostgreSQL and SQLite, including 29 Chrome tests covering keyboard use, phone widths, accessibility and judge access without losing an open order. These verify software behaviour, not visual accuracy.
What we learned
A useful exchange needs eligible choices, a clear price, visible uncertainty and control over what changes. The demo uses fictional commerce data, generated people and genuine recorded YouCam outputs. It changes no real order, stock or payment. We have no independent human ratings or shopper/merchant outcome study, and do not claim reduced returns or retained revenue.
What's next for Rechoice
Collect the registered blind human labels, observe shoppers completing an exchange decision, then pilot with a store's real catalog and policy. Measure whether shoppers understand the warnings and whether staff find the requests useful before claiming business impact.
Built With
- anthropic
- axe-core
- celery
- claude
- css
- django
- docker
- html
- javascript
- pillow
- playwright
- postgresql
- python
- redis
- youcam-api