Gen-AI Try-On · Consumer / B2C · 2026
Shipped a gen-AI try-on app to 250+ users, then found out why they would not upload a selfie
- ROLE
- Sole designer. Product design and front-end build.
- TEAM
- One full-stack developer, one AI engineer, me. No product manager.
- COMPANY TYPE
- Startup, Consumer B2C
- DELIVERABLES
- Research, Product Design, Branding, Design Language, Design System
- MEDIUM
- Mobile App, Website
TOOLS
- Figma
- ChatGPT
- Claude
- VS Code
- Vue.js
- Tailwind
- Next.js
- Fashn API
- MongoDB
- AWS
- Vercel
At a glance
- Problem
- Shoppers can’t tell how clothes will look on them online, so they buy several sizes and return most, or don’t buy.
- What I did
- Designed the product and built the front end, from a one-line idea to a live beta. Then redesigned onboarding when launch data showed people stopping at the selfie screen.
- Outcome
- 250+ users, and first-timers now pick a demo model and go on to try on an outfit instead of dropping off.
- 250+
- Users post MVP launch
- +80%
- First-time users choosing a demo model instead of dropping off at the selfie screen
- 60%
- Of first-time users completed a try-on in under 5 minutes
Challenge & context
Appari is a gen-AI virtual try-on app. I took it from a one-line idea to a live beta with 250+ users, designing the product and building the front end alongside a two-person engineering team.
Then the launch data showed something the MVP metrics had hidden. People were signing up, reaching the screen that asked for a photo of themselves, and stopping.
The second half of this case study is what I did about that.
A gen-AI virtual try-on app I took from a one-line idea to a 250+ user beta. Then launch data showed people stopping at the selfie screen, so I redesigned onboarding.
Proof of concept does not exist yet
You can't tell how clothes will look on you online, so people buy multiple sizes and return them, or don't buy at all. We all have experienced this.
The bet: show shoppers clothes on their own body, from any store, and they'll buy with more confidence. A real zero > one.
What 340+ people told us
From surveys and direct interviews, main signals:
“I usually buy two sizes, try them on, and return one.”
“If I could see how it looks before buying, I'd shop online more often.”
“I don't look like the model, I can't picture myself wearing it.”
Competitors
Competitors
Appari
No social integration
Social sharing built in
Complex setup
One full-body photo
Poor mobile experience
Mobile-first app
Single store only
Any retailer, from a screenshot or saved image
How I worked
- Research
- Drafted survey questions, then analysed the response data into the core problem themes.
- Design review
- Without co-designers on the team, ran mock-ups past a UX GPT to find holes before visual prototyping.
- Coding with Codex
- Used Codex connected to Visual Studio to code front end components and style library from a design POV and visual polish.
- Manual override
- Ignored many AI suggestions where it was suggesting features based on user research it had hallucinated. I always get it to double check facts and information.
What I did
Part one
Getting it built
Zero to a live beta
01One selfie, not a body scan Instant
Cut face and body scanning down to one full-body photo; onboarding dropped to under a minute.
02Dropped Body Measurements
Left sizing out of the MVP: the AI couldn’t predict fit reliably yet.
03Removed Outfit URL
Removed outfit-from-a-link for the MVP: the tech couldn’t generate an outfit from a URL in time.
04Simplify Onboarding
Swapped full account creation for ‘Login with Google’, so people could try the app sooner.
How it works
What we started with. The core user journey: import an item from any retailer, upload a selfie or saved photo, see it modelled on their AI twin, then share, or proceed to purchase.
I prioritised features for the initial launch based on user needs and technical feasibility:
- Basic try-on with photo upload
- Cross-retailer functionality
- Mobile-optimised interface
- Social sharing capability
- Freemium credit system
- User AI Model
Throw away wireframes for a new behaviour
Because there isn't a UI kit, or existing app, I explored this using quick throw away wireframes for initial ideas.
Create AI model. To create the most accurate digital twin: scan face & body with camera or upload images, then input body measurements for accuracy.
Feedback: “Too long winded, too many steps. Needs to be simpler.”
Try on clothes. Option to input link or screenshot/image of the items. I added images/screenshots as clothes can be captured from influencers. Clothing is remodelled on the user's digital twin.
Feedback: “This is simple and straight forward. For tech reasons, we couldn't generate the outfit from the link.”
01One selfie, not a body scan Instant
Cut face and body scanning down to one full-body photo; onboarding dropped to under a minute.
- Problem
- The first concept: scan your face and body, add measurements. Users said it plainly: too many steps.
- Tried
- Face and body scanBody measurements
- Shipped
- One full-body photo.
- Why
- The AI worked well from one full-body photo, so I cut the scanning.
- Result
- Onboarding dropped to under a minute.
Accuracy nobody experiences is worth less than a first try-on everybody finishes.
02Dropped Body Measurements
Left sizing out of the MVP: the AI couldn’t predict fit reliably yet.
- Problem
- Users told us sizing was a huge issue and the main reason for returns.
- Tried
- Body measurements to predict size
- Shipped
- Left sizing out of the MVP.
- Why
- The try-on tech of the time needed long research and training to predict sizing accurately. It didn’t exist yet.
03Removed Outfit URL
Removed outfit-from-a-link for the MVP: the tech couldn’t generate an outfit from a URL in time.
- Problem
- People usually save or screenshot clothes they like, and share them with friends. The design took a screenshot or a product URL.
- Tried
- Generate the outfit from a product URL
- Shipped
- Outfits upload from the user's image library, or as screenshots from social media.
- Why
- Tech limits and time: we couldn’t generate the outfit from a URL, and wanted a quick solution for the MVP.
04Simplify Onboarding
Swapped full account creation for ‘Login with Google’, so people could try the app sooner.
- Problem
- Full account creation, with email and password, took four screens.
- Tried
- Email and password sign-up, 4 screens
- Shipped
- ‘Login with Google’, 1 screen.
- Why
- For the MVP we needed to onboard people quickly, get them trying the app, and cut friction.
What shipped
The live app, a hand-coded Tailwind design system, and the marketing website.
They signed up, reached the screen that asked for a photo of themselves, and stopped.
What I did
Part two
Getting it used
What the launch data showed
05Models before selfies Trust
Added demo models under the selfie upload, so people see a result before sharing a photo of themselves.
Modified, demo models added
06Demo clothes, no blank slate Instant
Pre-selected demo clothes, so a first try-on never depends on having an outfit photo ready.
Modified, demo clothes added
What the data said
The MVP numbers looked like a finish line. They were only counting sign-ups.
Prototype testing, 4 users
- Tested Figma prototype with key user flows
- Identified confusion around photo upload requirements
- Additional feedback from UX GPT
- Issues around credit system and subscriptions
Beta launch, 87 users
- Real product testing with live AI generation
- Strong validation for core concept
- Feedback on load times and quality expectations
- Errors
- Trust issues around body selfie
Talking to the people who left
“People weren't put off by the technology. They were reluctant to hand a photo of their body to something they had no reason to trust yet.”
Mixpanel showed people abandoning at the selfie screen, the first step of onboarding, so they never tried on an outfit. Until then the only metric was sign-ups, which counts users, not value. The new north star: a user tries on two outfits and sees the value.
I emailed everyone who didn't upload a selfie or try on an outfit, and watched people use the app in person. Two reasons came back:
Reason 1
Trust with uploading a selfie
Uploading a full-body photo to a new app felt vulnerable. Users didn’t know where their photo was going or how it would be used.
Reason 2
No clothes photo ready
The blank slate problem. First-time users had nothing to try on yet, so even with a photo uploaded, the next step felt empty.
Three worries, one design goal
Three worries stood between a curious user and a first try-on.
Transparency & data privacy
“Where do my images go?”
A full-body selfie is a vulnerable ask. Offer demo models as a way in, and never charge for demos.
AI slop
“If it fails, I lose credits.”
A bad generation is not the user’s fault. Re-try is unlimited, at no charge.
Social proof & credibility
“This is a new way to shop.”
Social sharing builds trust, because it comes from friends.
Design goal. North star: try on two outfits, feel the value.
Each idea: idea → design → test with market → valid or invalid → iterate
05Models before selfies Trust
Added demo models under the selfie upload, so people see a result before sharing a photo of themselves.
- Problem
- The MVP asked for a full-body selfie straight away. Users told us that felt vulnerable on a new app.
- Shipped
- A “No selfie? We got you” option right under the upload, with a range of demo models.
- Why
- People see a result first, with no photo of themselves. Demos never cost credits.
- Trade-off
- It’s better to see clothes on yourself.
Modified, demo models added
06Demo clothes, no blank slate Instant
Pre-selected demo clothes, so a first try-on never depends on having an outfit photo ready.
- Problem
- First-time users had nothing to try on, so even after uploading a selfie the next step felt empty.
- Shipped
- Pre-selected demo clothes.
- Why
- A first try-on never depends on having an outfit photo ready.
- Trade-off
- It’s better to use clothes you like.
- Note
- The redesign mock also shows a paste-a-link field. That was a design exploration and is not live in the app.
Modified, demo clothes added
Impact & evidence
250+
Users post MVP launch
60%
Of first-time users completed a try-on in under 5 minutes
+80%
First-time users choosing a demo model instead of dropping off at the selfie screen
90%
Of users who started with a demo model went on to try on an outfit
2 outfits
Most users reached the north-star metric, using models and demo clothes
15%
Chance of a generation miss. Retries are unlimited and free.
Behaviour followed the same order. People tried their own outfits on a model by their third try, then moved to a full selfie upload and their own clothes.
What changed after launch
What we saw
What changed
Users hesitant to upload personal photos immediately
Demo models and outfits to explore first
AI generations not always to a good standard
Regenerate, without charging another credit
On first use, people didn't have clothes to try
A gallery of clothing to test with
Credits too expensive
Lower pricing and more free credits
What I learned
- For a behaviour nobody has tried before, adoption beats accuracy. The body scan and measurements would have given better results, but people wouldn’t finish them.
- Launch showed a second layer: even one selfie was too much to ask first. Trust had to come before the photo, with a demo model people could try at no cost.
Would do differently
- Test feasibility at the wireframe stage. I designed the body scan and the paste-a-product-link flow before checking they could be built.
- Body and face scanning wasn’t realistic for this concept at the time, and would have meant long onboarding for something nobody had tried before.