Project Done at AI Verticle of PagarBook
What is the Product
Flyr is a mobile app that generates ad creative from a single product photo. Small sellers who can't afford a photographer upload one phone photo and get images and short videos of their product on a model, built around the occasion they're selling into.
Business Bet
In November 2025, one generated product image cost us ₹1–2.50 plus GST. Eighteen months earlier it took a photographer, a model, a studio and a day. The same tooling made the platform cheap to build .We picked the buyer before we picked the problem. The filter was willingness to pay, not depth of need — plenty of people want free AI images; almost nobody pays ₹700 a month for them.
Choosing the buyer before we chose the problem
Casual consumer, tier 3–4
Has Need of AI Generated Images
Will not Pay for it
No revenue attached to the output, free alternatives everywhere, images are a novelty, not an quirk.
Small business, 3–4 people
Has Need of AI Generated Images
Will Pay for it
Output feeds directly into sales, already spends on photography, has expendable income
Key Questions that needed Answers
Q1: Where do they sell and does every channel want the same thing?
Q2: How does a product get photographed today?
Q3: What happens when the seller is the one doing the selling?
Q4: Can they tell a good image from a bad one and can they say why?
All four needed the same thing: going and looking. So that's where I started — with the listings, and then with the people behind them.
The Observation
I was scrolling Meesho, looking at kurtas under ₹500. Every seller had the same three photos: the garment flat on a bed, the same garment slightly crumpled, and a close-up of the fabric. Different sellers, same three photos. Forty listings in, I couldn't tell any of them apart — and neither could a buyer.
These sellers aren't careless. Most are two or three people running a business out of one room. A basic product shoot costs ₹200 per SKU; one with a model costs ₹500. They add 3-4 new SKUs a month. The arithmetic has never worked for them, so they shoot on a bedsheet with a phone and move on.
Platform Problem
Then I found the thing that didn't fit. The same sellers were moving the same products, at the same price, over WhatsApp and moving them well. Same photos. Same product. One channel worked; the other didn't. So the photos weren't the whole problem. Something else was carrying the sale on WhatsApp and going missing on the marketplace.
How Business was ran
Flyr's seller is two or three people, often one family, running a business out of a single room in a tier 2–4 city. They add four or five new SKUs a month not many, which means each one has to earn its place. And they list the same stock across Meesho, Amazon and Myntra.
Those three aren't interchangeable. They're a ladder, and photography is what gates it. Meesho will take almost any image. Amazon has standards. Myntra has real requirements name them: model shots, background specs and a seller who can't produce that imagery doesn't get listed there at all, however good the product is.
So the cost of bad photography isn't only lost sales on the listings they have. It's being locked out of the channel where their margin would be best.
Insights from the User Research
Photography isn't how they compete on a marketplace. It's how they qualify for one.
Output could never be "a product photo". It had to be whatever each rung of the ladder demands which is where the four formats came from, including the branded status creative that no advertising tool builds.
There is no second photo. What they have is all they will ever have.
One photo in, permanently. Every idea that quietly depended on "take a better one" — guided capture, retry prompts, quality gates — came off the table in the same week.
They don't sell the product. They sell the certainty around it.
The image had to carry context, not just quality. An occasion, a person wearing it, a sense of where it belongs. "Make the photo sharper" stopped being a plausible answer.
They know exactly what they want. They just can't say it first.
The direction of the whole interaction. Asking them to describe wants the one thing they can't give; showing them something to react to asks for the thing they're best at. Choose an outcome, don't configure an input — and every decision in Act 2 is an application of that.
What I was solving for wasn't better Images.
It was getting a seller who can't describe what good looks like to a publishable creative across platforms.
What design had to survive
constrains the users brought
Digital Literacy of users
Sellers skim. Many read English slowly, some barely at all.
Design Vocabulary
"Lifestyle", "editorial" mean nothing. They can describe the image.
Need of customization
Giving control of image feels like they control output.
Units of Products
Four or five new SKUs a month, worked on in bursts between orders.
constrains the System brought
Systematic Prompt Setup
Making the prompt lineup such that the customization fall in line rather than destroying that intent of product image of the users.
Wrong image costs more
A creative showing detail the product doesn't have gets ordered, returned, and the seller eats the shipping and the rating.
Generation isn't instant
A server round trip of 8 seconds, on a budget Android over patchy 4G. A dropped connection mid-generation is normal, not an edge case.
Problems to Solve
One photo in, One photo out
Three decisions, not Thirty
Nothing to read
No vocabulary to learn
Faithful, not flattering photos
Right the first time
Works between platforms
Introduce trust elements
Ideal Flow
1
Upload
2
Select Style
3
Customization
4
Generate
Compititor Analysis
What they share is an assumption about who's holding the phone. Background removal, AI shadow, recolour, product staging, brand kits, instant resize, batch export.
That's the correct product for a Shopify seller in Austin with some visual literacy and an afternoon. It is the wrong product for a two-person shop in Meerut adding four SKUs a month between orders not because the tools are weak, but because every one of them asks a question our seller can't answer.
They hand you tools. We had to hand over taste.
PhotoRoom gives a seller background removal, AI shadows, recolour and product staging, then trusts them to compose something good. Our seller has no reference for what good looks like.
They put the product on a background. We put it on a person.
The editor tools stage a product in a scene. For apparel and jewellery in India the model is the sale a kurta on a hanger and the same kurta worn are not the same listing, and Myntra won't take the first one at all.
They format for storefronts. We format for status.
Instant Resize covers Instagram, Amazon and Shopify. It doesn't cover the channel these sellers actually own WhatsApp status, posted several times a day to people who've already bought once which is why Flyr sets the shop's branding into the image itself.
What success would look like
High publish rate without editing
High user retention rate till M3
Creatives end up on live listings
Decision 1 - What a Shoot Actually Produces
We opened with two categories — clothing and jewellery because those were where the research said the gap between what a seller could shoot and what they needed was widest. Everything else came after: electronics, accessories, art and craft, kids, food. Add the order and rough dates. Category isn't a label on the output. It changes what the system does.
Image
Video
Catalogue
Branding
Decision 2 - What Level of Customization
Generation models expose dozens of parameters. Lighting, lens, pose, colour grade, aspect ratio, camera angle. Every one is a decision the seller has no basis to make, and every one is a chance to produce something worse than the default.
The obvious move is to strip all of it out and fully automate. We tried. Sellers rejected it. It's their product and their customer, and handing over the entire decision felt like handing over the shop.
Designs Iterations
Home screen — three directions and a refinement
Present it as exactly that. Screens 1–3 are genuinely different organising principles; 4 is a refinement of 3. Calling a header removal a fourth concept is the kind of thing a reviewer notices, and labelling it honestly costs you nothing.
Creation Page
Eighteen sellers, moderated sessions, in person / remote.
Everyone brought their own product. Not a sample photo, not a stock SKU a thing they were actually trying to sell that week. That was the only part of the setup I was strict about, because control choices are meaningless without a real buyer in mind. Asked to pick a model for someone else's kurta, a seller shrugs. Asked to pick one for their own, they have an opinion immediately.
Control
Model Face
Special instructions
Aspect Ratio
Busines Details
Cloting Settings
Of 18 USers
14 changed it
9 used it
7 changed it once
5 touched it once
3 changed it once
Paterrn
Repeatedly, and asked for it upfront
Wanted to add some details
Set once, never revisited
Tick-box. Set it, forget it
Did'nt care
Creation Page Iterations
Image Generation Flow
Where I was wrong about the market
I was confident clothing would be the entry point and the revenue driver. It's the biggest category, the most visual, and the one where bad photography costs the most.
Research disagreed. Jewellery came back just as strong, for a reason I hadn't considered: a jewellery seller's product is small, reflective, and almost impossible to shoot well on a phone, so the gap between what they could produce and what they needed was wider than in apparel.
%
Cloting Category
%
Jwellary
%
Rest
Impact
Phase 1 · Dec 2025 – Jan 2026 · Photoshoot + Branding
K +per day
Images Generated
%
Retention till M3
%
LTP
Phase 2 · Feb – Jun 2026 · Catalogue added
K +per day
Images Generated
%
Retention till M3
%
LTP
Phase 3 · Jul 2026 – present · Video added
K +per day
Images Generated
%
Retention till M3
%
LTP
A Feature No one asked
Open Instagram, search jewellery, and you get a wall of it. Nobody teaches you how to use that screen. You scroll until something stops you, and then you know what you want.
So the feature currently in testing is that screen, inside Flyr. A seller opens their category and gets a wall of finished creative Indian models, international models, studio, street, festive. They tap the one they like. They upload their product. They get their product, in that pose, in that scene.
