OnFoot Case Study
OnFoot
UI/UX case study · B2B SaaS · 2026

OnFoot

Traditional footwear shoots were slow and costly for catalogues of 5,000+ SKUs and frequent releases. Generic AI looked faster, but repeated generations wasted tokens and hallucinated shoe details. OnFoot made on-model imagery scalable around one rule: change the model, outfit and scene, never the shoe.

RoleLead product designer
ScopeResearch, UX strategy, UI, design system
Timeline9 weeks to beta
MarketFootwear retail & brands
Team7 people · 3 design partners
app.onfootai.com
OnFoot Studio with a generated take of a brown leather shoe, a strip of scored takes and the shot brief panel
OnFoot on a phone
Executive summary

Deloitte’s Seeing Is Believing report on AI-powered virtual styling sparked an idea. I explored the market and found footwear was largely overlooked. Talking with manufacturers and graphic designers revealed costly photoshoots, wasted AI generations, and hallucinated shoe details. I developed OnFoot AI to solve this, cutting production from 3–4 weeks to under a minute, at $0.003 per generation versus $80–150 for photography, with an 87% average shoe match. Guided briefs reduced retries, while accuracy checks caught AI hallucinations.

<1 minto on-model, from 3–4 weeks
$0.003per image, from $80–150
87%average shoe match
88%less AI spend per usable image

The bet

Focus on the overlooked footwear niche: reduce wasted generations, make AI product errors visible to designers, and deliver accurate on-model imagery at scale.

My role

Lead product designer. Research, strategy, interaction model, design system, and the case that changed the roadmap.

Read in 2 minutes

How I changed minds · Decision log · AI trust · What I'd do differently

The short version
Challenge

A Deloitte report inspired me to explore practical AI value. I identified footwear: expensive shoots, vast catalogues and designers losing time and tokens as generic AI hallucinated shoe details.

Solution

What began as a solution for one manufacturer became footwear-specific SaaS: guided briefs instead of long prompts, two takes per run, internal shoe-match checks, plus shared workspaces, usage controls and approvals.

My role

Lead product designer on a cross-functional team of seven. I led the market and user research, UX strategy, the interaction model and the design system, from discovery to beta.

Impact (beta, 8 weeks)
87%average shoe match
$0.003per image, from $80–150
6 minto on-model, from 3–4 weeks
88%less AI spend per usable image

OnFoot turns a brand's existing product photos into on-model photography: a real-looking person wearing the exact shoe, in a chosen outfit and scene, sized for the channel it's going to.

Footwear brands rely on packshots because model shoots are expensive and slow. Market research showed AI on-model tools mostly served clothing. Manufacturer conversations confirmed the gap: generic AI made attractive scenes but changed shoe details, and repeated prompts consumed tokens without reliable results.

I led design for OnFoot from the first market scan to a paid beta, working with product management, engineering, ML and QA, and with three footwear companies as design partners. This case study walks through what the research found, the decisions behind the product, and why each one was made.

Process

Nine weeks from first interview to paid beta

I worked through research, design and build side by side. I tested concepts against design partners' products and real model outputs each week, so my designs stayed grounded in what the AI could deliver.

Role & team

A product team, with clear ownership

OnFoot was built by a cross-functional product team of seven, with three footwear companies as design partners. I led design and research; decisions were made together and written down with an owner.

AS

Ali Syed Tabish

Lead product designer
  • Led the market scan, 18 interviews and the AI tool and spend audit
  • Defined the UX strategy, brief model, IA and flows
  • Designed every screen and owned the design system
  • Co-designed the generation and accuracy-check prompts with ML
  • Planned and moderated three usability rounds
PT

Product & engineering

PM · 3 engineers · ML engineer · QA
  • PM: pricing, packaging and the beta programme
  • Engineering: API layer, workspaces, SSO, storage and usage metering
  • ML: image pipeline and tuning the match check
  • QA: test plans for every state and edge case
DP

Design partners

Three footwear companies
  • A multi-brand footwear retailer with 5,000+ SKUs
  • A DTC footwear brand releasing every 2–3 weeks
  • A footwear marketplace seller group across 3 regions
  • Real catalogues to test on, and weekly reviews of takes
Daily 20 minDesign, product and engineering sync on what shipped yesterday
Weekly FridayDesign-partner review on their own SKUs, no slides
Decision logEvery trade-off written down with the reason and owner
One shared libraryDesign tokens defined once and used in both Figma and code
02Chapter 02 of 06

The problem

Footwear emerged as the gap: apparel tools prioritized clothing, while manufacturers needed their actual shoes preserved.

  1. 02.1Market research
  2. 02.2The problem
  3. 02.3Problem deep dive
Market research

How I found footwear

I began with a Deloitte report on the gap between AI experiments and business value. My scan of the on-model imagery market found plenty of tools for clothing, but little built around footwear accuracy. I contacted manufacturers to test whether that gap was real.

Apparel-first on-model toolsBuilt around garments: flat-lay to model, try-on, model swaps
BotikaClaidFASHN AIModeliaWearViewUwear

Strong for clothing. Shoes are treated as part of an outfit, not the product being sold.

E-commerce editors with AI modelsGeneral product-photo tools with a fashion-model feature
PhotoroomPixelcutAyna

Fast and cheap. Footwear, where listed, is one input among bags and jewellery, with no product-accuracy check.

General AI image and LLM toolsWhat design teams were actually using day to day
ChatGPT imagesGeminiMidjourney

Flexible, but every image starts from a prompt, every retry costs tokens, and nothing checks the shoe.

The gapNothing in the market was
  • Built for footwear first
  • Checking the product on every image
  • Made for catalogue scale and frequent releases
  • Controlling AI spend across a team

Tools grouped by their public product positioning, October 2026. Test results below are reported by category, not by vendor.

Demand signals
5,000+SKUs in a typical catalogue at the retailers we spoke to, each in several colourways
2–3 wksbetween releases at fast-moving brands, against 3–4 weeks to shoot on model
7 of 9footwear companies had tried an apparel AI tool on footwear, and dropped it
$14.6ka month one retailer's design team spent on generic AI subscriptions and tokens
How I researched it
  1. 01

    Market scan

    Mapped 14 AI on-model and image tools by category, input type, pricing and what they promise about the product.

    2 days · desk research
  2. 02

    Footwear stress test

    Ran 60 real SKUs from the design partners through each category, scoring whether a buyer would notice the shoe had changed.

    1,080 generations
  3. 03

    Interviews

    18 people at 9 footwear companies: e-commerce, merchandising, creative and design leads, and the people who own the budget.

    45–60 min each
  4. 04

    Spend audit

    Reviewed three months of AI subscription and token invoices from four design teams, matched against how many images were actually used.

    4 teams · 3 months
What I found
71%

Apparel tools hallucinated shoe details

Shoes came back with a different toe, sole or stitching in most outputs. The tools were tuned for drape and fit of fabric, not the structure of a shoe.

11×

Retries per usable image

Designers using generic LLM image tools rewrote prompts about 11 times per usable image. Every retry was paid for.

82%

Of paid generations thrown away

Across the four teams, most of the AI spend went on images nobody could use, spread across personal accounts nobody could see.

6%

Of the catalogue shown on a model

At 5,000+ SKUs and a release every few weeks, shoots only ever covered the hero products.

Why I focused here

Manufacturer conversations made the opportunity concrete: designers were spending tokens on repeated prompts and still getting the wrong shoe. Footwear needed a repeatable workflow, measurable accuracy and predictable cost—not another generic image tool.

The problem

5,000 SKUs, a release every few weeks, and no way to show them worn

Online footwear listings often rely on packshots: a shoe on white. On-model imagery shows how a real product looks with clothing, but traditional shoots are too slow and expensive to cover every release. Generic AI offered a shortcut, yet it changed shoe details and consumed tokens through repeated attempts. Designers needed a reliable way to produce and verify the image before publishing.

Problem statement

Footwear businesses need on-model images for 5,000+ SKUs and every new release, but shoots cover a fraction of it, and the generic AI tools their designers use change the product and waste tokens on retries, so most listings ship with packshots only.

How might we

How might I help footwear brands put every product on a model in minutes, preserve shoe accuracy and make AI spending predictable?

The core problem

Generic AI hallucinates shoe details

Generic AI can make a convincing picture while hallucinating product details: distorting the toe, stitching, sole or materials. Longer prompt-and-retry sessions made it harder for designers to keep the product consistent. The person, outfit and scene could change; the shoe could not.

What the AI changed
  • Toe shape: round becomes pointed, or the reverse
  • Sole: colour, thickness, rubber instead of leather
  • Stitching and broguing: added, removed or simplified
  • Invented details: tassels, buckles, logos, extra eyelets
  • Material and colour: suede reads as leather, tan as brown
Why it had to stay exactly the same
  • The listing shows a product the customer won't receive
  • That drives returns and "not as pictured" complaints
  • Marketplaces require images of the actual item
  • A shoe you don't make, under your name, is a brand and legal risk
So OnFoot is built around one rule: change the model, the outfit and the scene, never the shoe.

Business goals

  • Trial to paid conversion ≥ 15%
  • Monthly retained brands ≥ 80%
  • Gross margin per take ≥ 85%
  • AI spend per usable image −80%

User goals

  • First usable image < 10 min
  • Prompt writing needed None
  • Confidence the shoe is right Every take

Success metrics

  • Average shoe match ≥ 8 / 10
  • Takes downloaded per run ≥ 1.5
  • System Usability Scale ≥ 80
Problem deep dive

Why the obvious fix never worked

“Use AI to make model photos” sounded like the obvious shortcut. But manufacturers described long prompt sessions, wasted tokens and shoes that no longer matched the original. I traced those failures back to the workflow before deciding what to design.

Symptom

94% of footwear listings in my audit lacked on-model imagery.Packshots show the shoe, but not how it looks when styled and worn.

Why 1

On-model photography costs too much to cover a catalogue.A shoot day covers 15–20 products; a 5,000-SKU catalogue with colourways needs more than 800 shoot days a year.

Why 2

Footwear multiplies variants faster than any other category.Every design × colourway × width × finish is a separate product, and new releases land every 2–3 weeks.

Why 3

AI should break the per-variant cost, but teams abandoned it or kept paying for nothing.7 of 9 companies had tried apparel AI tools and dropped them; design teams kept generic LLM tools and burned tokens on retries.

Why 4

The AI hallucinated product details, and nobody could tell how often.Generators optimise for a plausible picture, not a faithful product. Checking every image by eye was slower than a shoot.

Root cause

There was no dependable way to verify AI-generated shoes at catalogue scale, or to start without writing prompts.Making an attractive image was easy. Preserving the real shoe and finishing the work were not.

Effects
Lower conversionListings without on-model images converted at roughly a third of those with them.
Avoidable returns"Looked different than expected" was the second most common return reason.
Slow launchesReleases every 2–3 weeks; photography 3–4 weeks behind, so products launched without it.
Brand riskAn AI image of a shoe you don't make is a misrepresentation, not a styling choice.
Most footwear products are sold online without being shown on a person
Variant explosionCost of photography scales with SKUs, not with revenue.
Apparel-first AITools built for clothing, and general generators, alter toe shape, broguing, soles and colour.
Token wasteGeneric AI needs prompt craft; every rewrite is a paid generation, spread across untracked accounts.
No verificationNo tool told you whether the generated product matched the real one.
Causes
$2M

To shoot one year

What full on-model coverage would cost a 5,000-SKU footwear retailer from my interviews.

5,000 SKUs × 3 colourways ÷ 18 per day × $2,400 per day
$175k

AI spend a year, mostly wasted

One design team's generic AI subscriptions and tokens, with 82% of generations never used.

23 designers · $14.6k a month · 3 months of invoices
2.9×

Conversion gap

Add-to-cart rate of listings with on-model images versus packshot-only, across audited brands.

240 listings · 30 brands · public review counts as proxy

Technical constraints

  • No fidelity guarantee. The image model can be steered but never forced to keep the product exact.
  • 10–20 s per image. Waiting has to feel like progress, not a freeze.
  • Cost per call. Every take and every check costs money, so runs must be deliberate.

Business constraints

  • 85% gross margin per take, including the extra check call.
  • Enterprise-ready. SSO, workspaces, seat and usage controls, with a self-serve trial.
  • Agencies and brands on the same product, with separate client data.

Trust & legal constraints

  • Marketplace rules require images to represent the actual item.
  • Synthetic people need no model releases, but must not resemble real individuals.
  • Product photos are confidential before launch; they can't train shared models.
ProblemE-commerce managerDesign teamHead of digitalShopper
Cost of coverageBudget gone by springNot their budgetShoot spendNot visible
AI hallucinationsBrand riskWasted workLegal & returnsReturns
Token wasteSlow outputEndless retriesUntracked spendNot visible
Channel sizingDaily choreManual cropsConsistencyBad crops

Severity by role. AI product hallucination is the only problem every party feels, which made it the design centre.

03Chapter 03 of 06

Research

The problem was never just making an image. It was preserving the actual shoe while keeping retries and costs under control.

  1. 03.1Research findings
  2. 03.2Research synthesis
  3. 03.3Riskiest assumptions
  4. 03.4Who uses it
  5. 03.5Reframing the problem
Research findings

A convincing AI image was not always the actual shoe.

I spoke with the designers who make listings and the people who pay for them, then checked the findings against real AI outputs using the design partners' products.

18interviews across 9 footwear companies: e-commerce, creative, design and budget owners
240competitor product listings audited across 30 footwear brands
1,080generations across 14 tools using 60 real SKUs
4design teams' AI subscription and token invoices, three months each
Product accuracy 17/18 Token waste & retries 14/18 Release pace 12/18 Channel sizes 10/18 No approval step 8/18
AI hallucinations

"The model looks great. That isn't our shoe."

AI tools hallucinated product details in most outputs. Toe shape, broguing, welt and sole were the usual losses.

68%of AI outputs had visible shoe-detail hallucinations
Prompt fatigue

Teams don't have time to learn prompting

One usable image took long trial and error. Most sellers stopped after a few tries and went back to packshots.

11prompt rewrites on average before one usable image
Wrong shape

Good images broke when resized for a channel

Square crops cut the feet off portrait shots. Stories needed 9:16 and nobody generated for it.

54%of usable images lost the shoe when cropped for a second channel
Option teams usedProduct accuracySpeedCost per imageNo promptingChannel-readyReview step
Model photoshootExact3–4 weeks$80–150YesManualPhotographer
General AI & LLM image tools
ChatGPT, Gemini, Midjourney
Often changedSecondsCents × 11 retriesNoNoNone
Apparel AI model tools
Botika, Claid, FASHN, Modelia
Clothing-firstMinutes$1–3PartlySomeNone
E-commerce photo editors
Photoroom, Pixelcut
UnchangedSecondsCentsYesYesNo model
OnFootChecked every take< 1 minute~$0.20YesBuilt inScored

Competitive landscape by category, from public positioning and my footwear stress test. The gap: nothing combined a real model with a check on the product, at a cost a business could control.

Key insight

Teams needed a dependable image workflow: preserve the real shoe, avoid repeated prompting, flag inaccurate takes and make generation costs predictable.

Research synthesis

236 notes, five clusters, one pattern

I coded every interview into single-observation notes and sorted them in FigJam with the PM and engineering lead, without naming groups first. Five clusters came out. Bar length shows how many of the 18 participants raised each one.

The product must not change

71 notes · 17 of 18
"It gave my loafer a tassel. We don't make a tassel loafer."
"If the sole is wrong, the customer returns it. That's on us."
Checks every AI image at 200% before using it
Fidelity is the requirement, not a feature

We pay for every bad attempt

52 notes · 14 of 18
"We're on our third AI subscription and still booking shoots."
Rewrote the same prompt 9 times for one shot
"Nobody knows what we spend on AI. It's on twenty cards."
Replace prompts with choices; make cost visible

A catalogue has to look like one brand

44 notes · 13 of 18
"Every image looks like a different photographer."
Keeps a Notion page of "our look" nobody can reproduce
"Same model, same light, 400 shoes a drop. That's the job."
Saveable presets and a house style

Every channel wants a different crop

38 notes · 10 of 18
Crops one image three ways, loses the heel each time
"Amazon is square, Instagram is 4:5, the site is wide."
"I make the shoe tiny to fit the frame."
Frame and ratio belong in the brief

Someone has to sign it off

31 notes · 8 of 18
"I need to show my boss why this one, not that one."
Approvals happen in Slack screenshots
"Who made this, and with what settings?"
History with the brief attached, then approvals

Across the clusters, designers did not need prettier AI images. They needed a dependable way to spot changed shoe details, stop unnecessary retries and approve usable work.

Riskiest assumptions

What had to be true, and what we tested first

Before designing the UI, I listed the assumptions the product depended on and tested the riskiest ones with low-cost experiments. Two proved wrong, and I changed the design accordingly.

AssumptionRiskHow I tested itWhat I learnedStatus
A better image model will prevent hallucinationsRan the same 30 shoes through four models with tuned promptsBest model still hallucinated shoe details in 1 of 3 images. Model quality alone can't close the gap.Invalidated
Teams will review every image by eyeGave 6 people 40 images each, timed and scored their reviewsAccuracy fell to 61% after 15 images. People stop looking closely. Review has to be automated.Invalidated
Multiple reference angles improve fidelity1 vs 3 reference photos, 60 shoes, blind human ratingAverage match rose from 6.4 to 8.1 out of 10. Asking for more photos is worth the friction.Validated
Non-experts can direct a shoot without promptsPaper brief with 6 choices vs free-text prompt, 8 peopleBrief users matched the target image 7 of 8 times; prompt users 3 of 8.Validated
Shoppers accept AI-generated modelsListing A/B with disclosed AI images on a design partner's siteNo drop in trust signals, add-to-cart up. Disclosure didn't hurt as long as the shoe was right.Partly
An AI can judge product match reliablyAI judge scores vs 3 human raters on 400 takesStrong agreement (r = 0.81) with a few known blind spots. Good enough to flag, not to auto-delete.Partly

The two invalidated assumptions are why OnFoot has an automatic match check instead of a better prompt box.

Who uses it

Many roles touch every on-model image

In an enterprise, on-model imagery isn't one person's job. It moves through content, merchandising, catalogue and marketing teams, and is governed by brand and legal. Each role brought its own problem, and OnFoot had to solve for all of them.

People who use it every day
  • Content production team(someDM Sans called creative studio or content operations): photo producers, retouchers and visual designers. They run the briefs and generate takes in bulk for each new release. They're the main daily users.
  • Footwear e-commerce merchandisersThey own the footwear pages online. They request on-model images for new items, pick takes and check that coverage is complete before launch.
  • Catalogue and item-setup specialistsThey attach images to product listings and make sure each one meets the site's image rules (size, background, number of images).
  • Private-label product teamsFor the retailer's own footwear brands, they can create on-model images from early samples, before a physical shoot is even possible.
  • Marketing and social teamsThey reuse approved takes for campaigns, emails and social posts in the sizes each channel needs.
People who approve and govern it
  • Brand or creative directorSets the house style (models, scenes, lighting) as saved presets and approves takes before they go live.
  • Legal and complianceChecks AI-image disclosure, that models don't look like real people, and marketplace rules.
  • Quality reviewersLook at low-scoring or flagged takes.
Job to be done

When a new shoe or colourway is ready, I want accurate on-model images for each channel, so I can publish without a new shoot or repeated AI retries.

Stage
Product readyDay 0
Plan imageryDay 1–5
Shoot or generateWeek 2–4
Edit & resizeWeek 4
List & sellWeek 5
Doing
Packshots taken in-house on white
Book photographer, model and location, or try an AI tool
Shoot a sample set; AI attempts discarded
Retouch and crop per channel by hand
Most products listed with packshots only
Feeling
Pain
Packshots don't show fit or styling
Shoots cost $2–4k a day; generic AI burns tokens on retries
AI hallucinates shoe details; a shoot covers 15–20 products
Crops cut the feet; manual work per channel
Shoppers ask "how does it look on?"
Opportunity
Start from the packshots they already have
Replace prompts with a brief
Check every take against the product
Export sized and cropped per channel
On-model images on every listing

Current-state journey for a footwear e-commerce team, built from interviews. Each pain became a design opportunity.

Reframing the problem

The question changed twice before the design did

Each round of research made the problem narrower and more useful. The final framing is what every later decision was measured against.

First framing · week 1
How might I generate on-model photos with AI?

Why it fell shortGenerating was already solved. Teams had the tools and had stopped using them.

Second framing · week 2
How might I make AI footwear photos more accurate?

Why it fell shortNo model is accurate every time. Chasing accuracy alone left teams checking by eye, which they couldn't sustain.

Final framing · week 3
How might teams know, for every image, whether the shoe is theirs, without needing AI skills?

What it unlockedVerification and access became the product. Generation became an ingredient.

04Chapter 04 of 06

Strategy & decisions

Swap the prompt for a simple brief, and have the tool say when it isn't sure.

  1. 04.1Design principles
  2. 04.2Information architecture
  3. 04.3Decision log
  4. 04.4Influence & alignment
  5. 04.5Designing for AI trust
Design principles

Three rules every screen had to keep

Each principle comes straight from a research finding, and each one maps to something you can point at in the product.

Compare view with the reference product photos beside the take

The product is the hero

Getting the shoe right matters more than a pretty scene. Up to four reference angles go into every request, the prompt locks the design, and Compare puts the real shoe beside every take.

Why

Product hallucinations were the top reason teams abandoned AI.

How

Multi-angle references, locked prompt, Compare and Loupe.

Shot brief panel with the brief written as a sentence

Brief, not prompt

Nobody has to write a prompt. Teams pick outfit, model, scene and framing, read the brief back as one sentence, and OnFoot writes the prompt.

Why

11 rewrites per usable image; teams gave up.

How

Visual choices, a readable sentence, saved presets.

A take flagged with a 7 out of 10 shoe match and the difference described

Check before it ships

Every take is scored against the real product and the difference is named in plain words, so a wrong shoe is caught before anyone downloads it.

Why

No tool had a review step; mistakes reached listings.

How

Automatic match score, colour bar, issue note.

Why a structured brief, not free prompting

Free prompting was costing tokens and breaking consistency

01 · It wasted tokens

Every free-text prompt was a guess. Designers rewrote prompts about 11 times per usable image, paying for retries even when the shoe drifted.

02 · Results weren't uniform

Different people wrote different prompts for the same shoe, so lighting, framing and model changed from image to image. A catalogue never looked like one brand.

03 · So I pre-built the prompt

Teams choose the outfit, model, scene, framing and channel through a guided brief. Those controlled choices reduce prompt rewriting and make the intended shoe and scene easier to verify.

Free prompting11generations per usable image · look varies by person
Structured brief1.3generations per usable image · one look across the catalogue
Information architecture

Two places to work, one place to configure

A card sort with eight participants put "make images" and "find images" far apart, and everything else under settings. The app keeps that shape: Studio, History and Settings, all scoped to a workspace.

Workspace · one brand or client
StudioProduct photos
Brief and presets
Takes and match scores
Refine and export
HistoryWall and table views
Search and filters
Stars, reuse brief
Full view and export
SettingsMatch check on or off
Auto-save
Usage this month
Security status
AccountWorkspaces and roles
Plan and billing
SSO on Enterprise
Sign out
Upload photos1–4 angles Set briefor pick a preset Generate1–4 takes Match checkscored 1–10 ≥ 8? Exportsized per channel, saved Refineadd a note, shoe stays locked yesno

The core loop. A refine sends the take back through the check, so nothing reaches export unscored.

Concept exploration

Three layouts, one survivor

I prototyped three directions at low fidelity and ran them with six users on the same task: "Create an Instagram image of this loafer on a model."

Rejected

A · Prompt box

One big text field with the image below. Fast for experts; every novice stalled at the empty field. Task success 2 of 6.

Rejected

B · Step wizard

Four steps, one choice each. Easy the first time, painful on the tenth run, because changing one setting meant walking back. 4 of 6.

Chosen

C · Studio canvas + brief

Canvas left, brief right, takes in a strip. Everything visible at once, and settings stick between runs. 6 of 6, fastest repeat runs.

Decision log

Every major call, and what it cost

Good decisions still give something up. I logged the options, the choice and the trade-off I accepted, so the team could revisit any decision with the reasoning intact.

Trust in the AI

What happens around the score
01

What to do with a low-scoring take

Flag with score and reason

Auto-delete / Hide / Flag and explain

WhyThe judge has blind spots; deleting would destroy good takes it misread

Trade-offUsers can still export a 6/10 take

02

Pass threshold

8 out of 10

7 / 8 / 9

WhyBest balance in calibration: 5% false passes, 9% false flags

Trade-offSome acceptable takes get flagged

03

Takes per run

User chooses, default 2

1 / fixed 4 / 1–4 user choice

WhyVariety beats retries, but cost must stay visible and controllable

Trade-offAn extra decision before every run

04

Reference photos

1–4, with "more angles, closer match"

1 required / 1–4, encouraged / 4 required

Why3+ angles lifted match 6.4 → 8.1, but requiring 4 blocked teams with a single packshot

Trade-offSingle-photo runs score lower

Directing the shoot

How people tell the AI what they want
05

How users direct a shoot

Brief panel + optional note

Free prompt / Step wizard / Brief panel

WhyNovices succeeded 6 of 6; experts kept a free-text note for nuance

Trade-offLess creative range than a raw prompt

06

How options are shown

Photo tiles for outfit, model, scene and framing

Text labels / Line icons / Real photographs

WhyPeople picked the look they wanted in one go instead of guessing what "editorial" meant

Trade-offMore images to license and load

Look and feel

How the product feels on screen
07

Canvas

The take fills the canvas, near-black chrome, with a Fit toggle

Image in a framed card / Graphite surround / Full-bleed image

WhyReviewers judged faster when the take was as big as the screen; a frame and grey margins made shoes look smaller than on the product page

Trade-offControls sit on top of the photo and need scrims to stay legible

08

Where decisions live

A copper gradient panel for the brief

Same dark panel as the canvas / Warm brief panel

WhySeparates "set up the shot" from "judge the shot" at a glance, and gives the product warmth without tinting the photo

Trade-offA second surface colour to maintain in every state

09

Sign-in

A full-bleed product photo with a slim form bar on the left

Split marketing page / One image + form bar

WhyShows the promise before any copy, and repeats the studio's layout so the app feels familiar on first load

Trade-offLess room to explain the product on that screen

Business model

How teams start and pay
10

Getting a trial

Account, then workspace, then the studio

Sales call only / Two-step self-serve sign-up

WhyFootwear teams wanted to test on their own products before talking to sales; the workspace step sets default channels and export sizes

Trade-offSome enterprise leads skip the sales conversation

11

Pricing unit

Takes per month

Seats / Takes / Images exported

WhyMaps to actual costs and how brands budget for shoots

Trade-offUsers ration exploration near the limit

Influence & alignment

Changing minds with evidence, not opinions

The hardest design work happened off screen. Four DM Sans, a stakeholder wanted something reasonable that would have hurt the product. Each time I brought data they could check, offered a way to keep what they cared about, and let the result decide.

Leadership"Ship a prompt box. Every competitor has one."
What I did
Pulled three weeks of logs from the AI tool audit: designers averaged 11 generations per usable image. Then ran a side-by-side with six designers, same shoe, prompt versus brief.
Outcome
The brief won on speed and consistency. The prompt survived as an optional art-direction note, so experts kept their nuance. Generations per usable image fell to 1.3.
ML lead"Hide low-scoring takes. Users shouldn't see failures."
What I did
Shared the calibration results: the judge wrongly flagged 9% of good takes. Auto-hiding would silently throw away usable images and teach users the tool was hiding things.
Outcome
Flag and explain instead, plus a "Score wrong?" button whose reports feed the next calibration round. Trust in the score went up, not down.
Sales"Price per seat. Buyers understand seats."
What I did
Mapped the actual cost per run and how brands budget for shoots. Seats priced a 5,000-SKU retailer and a 40-SKU brand the same, while generation costs differed by 100×.
Outcome
Takes per month, with the cost shown on the Generate button. Sales kept seats as an Enterprise add-on for SSO and audit.
Brand directors"AI will put our shoes on off-brand people."
What I did
Instead of arguing, I gave them control: saved presets for models, scenes and lighting, and a Viewer role that can review but not generate.
Outcome
Two of three design partners made a brand lead the preset owner. The objection turned into the approval flow now on the roadmap.
Designing for AI trust

Calibrated trust, not blind trust

I did not want designers to accept AI output blindly. The shoe-match score is an internal quality check: inspect high scores, investigate uncertain ones and reject inaccurate takes. Each pattern below supports that decision before an image is published.

Pattern
Where it lives in OnFoot
Why
Set expectationsBefore generating

The button says 2 imgs + 2 checks and the brief states the frame and ratio before anything runs.

Users who know what's coming judge results fairly. Surprise reads as failure.

Show confidenceOn every take

A shoe match score from 1 to 10, colour-coded at 7.5, on the image and in the takes strip.

One number turns "is this OK?" into a quick decision. It's about the shoe only, so it's honest about what it checks.

Explain, brieflyWhen a take scores low

A plain reason under the score: "toe is more pointed than the reference", "sole colour differs".

A score without a reason gets ignored. A reason tells the user exactly what to look at, or what to refine.

Keep users in controlRefine and history

Refining creates a new take beside the old one. Nothing is overwritten, nothing publishes itself.

Users experiment more when they know they can't lose a good result.

Make mistakes cheapAfter a bad take

One-click fixes like Fix shoe accuracy and Lower angle, written as instructions, not prompts.

If recovering costs one click, a bad take is a small detour, not a reason to churn.

Add friction where it mattersExport

Exporting a take below 7.5 to a marketplace asks for a confirmation that names the issue.

The one moment a mistake reaches a customer gets a speed bump. Nowhere else does.

Let users correct itScore feedback

"Score looks wrong?" on every take. Flags go to a review queue and into the judge calibration set.

Users see their judgment counts, and the judge gets better on the shoes it gets wrong.

Be open about AIExported files

Exports carry "AI-generated" metadata and alt text written from the brief, ready for marketplace disclosure rules.

Brands carry the risk of a misleading listing. Disclosure built in removes a question from their legal team.

7%of takes where users overrode the score, which showed they trusted it, but not blindly
3%of takes flagged "score looks wrong", all reviewed weekly
0below-threshold takes exported to a marketplace without a confirmation
4.4/5agreement with "I know when to trust a take" after a week of use
05Chapter 05 of 06

The product

An enterprise studio for judging shots at catalogue scale, from the landing page and sign-up to export.

  1. 05.1Design evolution
  2. 05.2The product, fully realized
  3. 05.3Mobile
  4. 05.4Onboarding & first run
  5. 05.5Accessibility & edge cases
  6. 05.6Marketing site
  7. 05.7Brand strategy & visual identity
  8. 05.8Design system
  9. 05.9Engineering collaboration & handoff
Design evolution

Three versions, and what each one taught me

These are real builds, not redrawn mockups. Each was put in front of users and design partners, and each failed in a specific way that shaped the next.

The original internal tool: a left column of form controls with emoji icons and a single result image
v0 · InheritedWeek 0

A form with a generate button

The first internal prototype. It worked for one person who already knew what to pick.

Failed: one result with no way to judge it. Users regenerated 6 times on average and still exported drifted shoes.
Second version: a shot brief written as a sentence, a strip of takes with match scores, and a serif editorial style
v2 · Concept BWeek 5

The brief as a sentence, takes with scores

The core model arrived here: a readable brief, two takes per run, a shoe match score on every take.

Kept: brief, takes, score, refine.
Failed: the editorial serif and dark craft look read as "a tool for one boutique". Two enterprise buyers asked whose product it was.
Current version: the take fills a near-black canvas, the brief sits in a copper panel, and photo tiles replace labels
v5 · ShippedWeek 9

A workspace product, image first

Same interaction model, rebuilt as SaaS: workspaces, seats and usage, export by channel. Visually, the take now fills the canvas, the chrome went near-black, the brief became one warm copper panel, and every option became a real photo.

Result: SUS rose from 61 (v0) to 74 (v2) to 86 (v4), and "whose tool is this?" stopped coming up.
The product, fully realized

A studio, not a prompt box

The take fills the canvas because judging the shot is most of the work. The brief sits on the right like an art director's sheet, in the one warm panel on screen, and takes collect in a strip along the bottom you can scan by colour.

OnFoot Studio, annotated
1 2 3 4 5 6 7 8 9
1

Workspace and usageAI spend used to go untracked. Brand and monthly usage stay in view.

2

Brief sentenceCatch mistakes before spending a run. Tap any word to change it.

3

Shot controlsNo camera jargon. Framing shown as photos, ratios named by channel.

4

Match badgeCatches a wrong shoe before download, and says what changed.

5

Takes stripSee which takes passed at a glance, without opening each one.

6

RefineThe notes people kept typing became one-click fixes.

7

ExportCrops kept cutting the shoe off. Each size crops toward the feet.

8

GenerateShows the cost in credits before the click, so spend is a choice.

9

Review toolsThe check is wrong about 5% of the time. Compare, Loupe, "Score wrong?" and Fit let people verify.

Generating state with a scanning animation over the frame
Generating: the frame takes the chosen ratio before the image exists, and each take shows its own progress.
Refined take added to the strip
Refined: the edit arrives as a new take (T4r), so the original is never lost.
Compare view
Compare slides the reference photos in beside the take.
Loupe zoom on the shoe
Loupe zooms into the shoe for perforations, welt and sole edge, and follows the cursor.
Workspace switcher showing three brands with roles
Workspaces separate brands or clients, each with its own roles, history and plan.
Sign-in: one full-bleed product photo with a slim copper form bar on the left
Sign-in: one full-bleed photo and a slim form bar, the same layout as the studio. Show/hide password, Caps Lock warning, clear errors.
History wall grouped by day
History groups takes by day at their real proportions. Star, reuse a brief, or open the full view.
Empty studio with a full-bleed photo and three numbered first steps
First load: a full-bleed example and three numbered steps, plus a sample shoe for people without a packshot to hand.
Full-screen view of a saved take
Full view of any saved take, with reuse-brief and download in channel sizes.
History table view
Table view for catalogue managers who search by SKU.
Settings drawer
Settings: workspace privacy, the match check switch and monthly usage.
Mobile

Brief on the floor, review on the go

Merchandisers and approvers review takes between meetings and on the shop floor, so on mobile the brief comes first, takes are easy to compare, and Generate stays pinned to the bottom of the screen.

Mobile sign-in with the photo above and the form below
Sign inThe photo stays, the form slides up.
Mobile brief with product photos and presets
Photos and presetShoot the product, tap a preset, generate.
Mobile canvas with takes
Takes and refineScored takes in a horizontal strip.
Mobile history wall
Workspace historyEverything the team has made.
Onboarding & first run

A good first take before the first question

For a self-serve SaaS, the first session decides the trial. The risk was an empty studio and a new user staring at twelve options. So the first run is guided, pre-filled and short, and it ends with a real result, not a tour.

Set up your workspace
Your company
Where do you sell?
Shopify
Amazon
Instagram
Continue
Step 1 · 20 sec

Tell us where images go

Channels set default ratios and presets, so the first brief already fits their store. Three fields, no team invite yet.

Add a product photo
No photo handy? Try a sample shoe
Step 2 · 30 sec

One photo, or a sample

A sample product means nobody hits a blank state. 31% of trials started with it, and 70% of them uploaded their own shoe next.

Your first brief
A male model, 30–40, smart casual, street, knee down, 4:5 for Instagram
Change anything, or keep it
Generate 2 takes
Step 3 · 10 sec

A brief written for them

The brief is pre-filled from their channel, as a sentence they can read. Every word is a control, but nothing has to change.

Shoe match 9/10
This score checks the shoe, not the photo. Below 7.5 we tell you what changed.
Step 4 · ~40 sec

Explain the score once

One tooltip, on the first take only, teaches the most important idea in the product. It's dismissed in one click and never comes back.

4 min 30 smedian time from sign-up to first exported take · target under 5 min
64%of trials export a take within 24 hours (activation) · from 22% with an empty studio
0required fields before the first generation, beyond the photo
Accessibility & edge cases

Designed for the bad days too

AI fails more often than ordinary software. Rate limits, blocked images and quota run-outs are everyday events, so each one has a designed message that says what happened and what to do next.

Contrast

  • Body text on graphite 15.8 : 1
  • Secondary text 7.6 : 1
  • Gold on graphite 8.9 : 1
  • Labels at 11px 4.9 : 1

Keyboard

  • Generate ⌘ / Ctrl + Enter
  • Refine Enter in field
  • History viewer ← → Esc
  • Visible focus 2px gold ring

Not colour alone

  • Match score Number + bar
  • Selected option Border + square
  • Failed take Icon + label
  • Screen readers Takes announced

Quota reached

Shown before the run, not after.

You've used 600 of 600 takes this month. Upgrade or wait until 1 November.

Take blocked

One failure doesn't stop the batch.

T3 failed: no image returned. Try another pose or scene.

Save failed

A take is never lost silently.

Not saved to History. Download it to keep it.

Busy service

Retries happen automatically.

Service busy, retrying in 3 s (attempt 2 of 3).
Marketing site

Sell the guarantee, not the AI

Brand buyers were tired of vague "AI magic" claims. The site leads with footwear imagery and clear production benefits; inside the product, designers see shoe-match scores and channel readiness before they approve an asset for use.

OnFoot marketing site hero: The AI photoshoot for footwear, over a full-bleed photo with a shoe-match chip
  • 1A full-bleed hero that rotates through dress shoes, heels, sneakers and boots. A frame locks onto the shoe, with chips for the match score and the channel it's ready for.
  • 2The real studio rises out of the hero, so the product is visible before any feature copy.
  • 3Benefits written for the buyer: fewer paid retries, one look across the catalogue, time back for designers, more coverage with the same team.
  • 4"Today" against "With OnFoot", a catalogue calculator, and an accuracy section with hotspots on toe, sole, stitching, details and colour.
  • 5One primary action throughout: start a free trial, which leads to the two-step sign-up below.
Full OnFoot marketing page
Sign-up

From the landing page to the first take in two steps

The trial is self-serve. Step one creates the account, step two sets up the workspace, so the studio opens with the right channels and export sizes already chosen.

Sign-up step one: work email, name and password over a full-bleed photo
Step 1: work email, name, password. A gentle note for personal email addresses, one-click typo fixes, and a live password strength bar.
Sign-up step two: company, team, catalogue size and channels
Step 2: company is pre-filled from the email domain; catalogue size and channels set the studio's defaults.
Sign-up confirmation with workspace summary and three first steps
Done: a summary of the workspace and the three steps to a first take, then straight into the studio.
Landing page on a phone
Landing on mobileShoe-cropped hero, menu, one action.
Sign-up on a phone
Sign-up on mobilePhoto on top, the form below.
Brand strategy & visual identity

A brand that frames the product, not competes with it

OnFoot's job is to make someone else's product look right. The identity is deliberately quiet: one mark, one typeface, a near-black workspace where the generated image fills the canvas, and one warm copper surface for everything you set up. The shoe is always the loudest thing on screen.

OnFoot

The frame mark

Four crop corners around a single shoe sole, tilted mid-step: a viewfinder with a foot in it. It says "on foot" without a word, reads at 16px and works in a single colour.

OnFoot — Studio
app.onfootai.com/studio
Aa

Geist · one family for the whole product

Light 300 · supporting copy
Regular 400 · body, inputs
Medium 500 · controls, labels
Semibold 600 · headings, brief
Display56 / 600 / −4%
Brief sentence20 / 400, choices underlined
Body14 / 400
Label11 / 500 / +14% caps
Ink#050506 · workspace and chrome. Near-black, so the photo is the only thing that glows.
CopperWarm gradient · the brief panel and sign-in bar. The one place you make decisions.
Brass#D4A843 · small accents only: the mark, plan badge, the take being refined.
Fog#F4F4F5 · text, and the one primary button per view.

The interface is a viewfinder. The take fills the canvas edge to edge, the chrome around it goes black, and the only warm surface is the brief, so your eye goes from the shoe to the decision and back.

Brand expression

Tight corners, thin dividers, real photography instead of icons, no italics, no pills. Precise and calm, like studio equipment.

Product context

OnFoot · AI on-model photography · B2B SaaS for footwear · 2026

Design system

Few components, every state designed

The app has about twenty components. What made it feel solid was designing every state of each one: hover, selected, disabled, loading and failed. These are the ones that carry the product.

Photo option tile

One component for outfit, model, scene and framing, each shown as a real photograph instead of a label or icon. Selection is a white border plus a check, which stays visible on any photo.

Full body
✓Knee down
Close-up
Studio
Street · hover
Disabled

Take thumbnail

Four states. The score bar reads at a glance, so nobody has to open a take to know whether it's usable.

Shooting
Checking
T1 · 9
Failed

Match scale

Semantic colour is used here and nowhere else, so green always means "safe to publish".

1
3
5
6
7
8
9
10
Product changedListing-ready

Buttons

One white primary per view, with the cost of the run written on it. Destructive actions are outlined red and always ask to confirm.

Generate 4 takesRefreshCopy imageClear all

Colour tokens

Four near-black surfaces, one warm gradient, one brass accent and three semantic colours.

ink · #050506
s1 · #0E0E10
s3 · #1E1E21
warm · gradient
text · #F4F4F5
accent · #D4A843
ok · #5FC48A

Spacing & radius

4px base grid. Corners are 6px on controls, 8px on cards and 10px on panels. An early round used 20px corners and capsule buttons; users called it "an app for consumers", so we tightened them. No pill shapes.

4
8
12
16
24
Engineering collaboration & handoff

Designing inside the cost and latency budget

Every generation costs money and takes time, so actual model capabilities shaped my design from the start. I set a budget in week 3 (under 45 seconds and under $0.25 per listing-ready take) and every interaction was checked against it.

Trade-offs I worked through

DecisionChosenRejected
Takes per run2 takes · 38 s · $0.104 takes · 71 s · $0.20. Users picked from the first two 80% of the time.
When the score appearsImage first, score streams in ~3 s laterWait for both. Felt 40% slower in testing for a 3 s gain.
Aspect ratioNative ratio from the modelGenerate square and crop. Cut off heels on 4:5 and 16:9.
RefineEdit the selected take with its briefRegenerate from scratch. Lost what the user liked about the take.
Model accessThrough a dedicated server, tied to each workspaceCalling the model straight from the browser. Faster to ship, but not safe for business customers.

Tokens, defined once

The same names in Figma variables and in CSS, so a design change is one edit.

InkPage surface · #0e0f11
Surface 1Cards and panels · #15171a
Surface 2Raised panels · #1b1e22
GoldAccent and main actions · #d4a843
PassShoe match passes · #2f9e63
FailShoe match fails · #cf4459
Score levelsPass at 7.5, warn at 6.0, the same numbers the match check uses
Spacing4-point grid · corner radius 4, 6 and 10

Spec: the take card

Written per component, with every state and the edge cases engineering asked about.

StatesGenerating · scoring · pass · warn · fail · error · selected
Score timingSkeleton chip until the judge returns; 12 s timeout shows "Not checked"
ErrorSafety block or failure keeps the brief and offers one retry; no credit charged
Keyboard← → between takes, Enter to select, R to refine, E to export
Screen reader"Take 2 of 2, shoe match 8 of 10, passes"
46handoff tickets, each linked to a frame and an acceptance check
23edge and error states designed before build, not after
2 roundsof design QA on the build, 31 issues fixed before beta
0P1 bugs at beta launch
06Chapter 06 of 06

Impact & learnings

Measured on takes good enough to sell with, not on images generated.

  1. 06.1Validating the check
  2. 06.2Usability testing & iteration
  3. 06.3Final results
  4. 06.4Measuring success
  5. 06.5From problem to outcome
  6. 06.6Learnings & next steps
  7. 06.7What I'd do differently
Validating the check

Can the AI be trusted to check its own work?

The whole product rests on the match score, so I tested it like a measuring instrument. Three footwear specialists rated 400 takes blind, and I compared their consensus with the AI judge.

Human consensus score AI judge score
86%of takes: judge and humans agree on pass or flag
r = 0.81correlation between judge and human scores
5%false passes: judge said 8+, humans identified hallucinated shoe details
9%false flags: judge was stricter than humans

Where it fails: colour shifts under warm scene lighting, and lace colour on dark shoes. Both are now named in the issue note so a human looks twice.

What it means for the design: the score flags, it never deletes. Compare and Loupe exist for the 5% the judge gets wrong.

Ongoing: 2% of production takes are sampled weekly for human review, so drift in the judge shows up before customers see it.

Usability testing & iteration

Three rounds, fourteen people, four changes that mattered

Moderated sessions on the working build, with participants' own products. Each round used the same three tasks so results could be compared.

TaskRound 1 · 5 peopleRound 2 · 5 peopleRound 3 · 4 people
Make a listing-ready image for Amazon
3 / 5
4 / 5
4 / 4
Spot the take with hallucinated shoe details
2 / 5
4 / 5
4 / 4
Fix one take and export for Instagram
2 / 5
3 / 5
4 / 4
Median time to first usable image9 min 40 s6 min 10 s3 min 50 s
System Usability Scale647786
Finding 1

A raw number didn't mean anything

Round 1 showed "Accuracy 0.74". Nobody knew if that was good, and two people downloaded a take where the toe had changed.

Before
Accuracy 0.74
After
Match 7/10
Toe reads rounder than the reference.

A score out of 10, a colour bar, and one sentence naming the difference.

Finding 2

Refine edited the wrong take

With five takes on screen, three people refined a different take from the one they meant.

Before
Refine…Which take?
After
REFINE T3 —lower camera angle

The refine bar names its target, and the result arrives as a new take.

Finding 3

Exports cut the shoes off

Centre crops for 1:1 removed the feet from portrait takes, which is the most important part of the image.

Before
Centre cropFeet cut
After
Amazon · 2000×2000Cropped around shoes

Crops are weighted to the lower third, and the ratio is chosen before generating.

Finding 4

People lost track of their settings

After a few changes, users couldn't recall what they'd asked for and re-checked every section.

Before
7 sectionsScroll to check
After
A male model, 30–40, in a formal suit…

The brief sentence reads the setup back. Time to verify settings fell from 22 s to 4 s.

Final results

From a pilot to a product brands pay for

Results from the eight-week beta, with three design partners and eleven footwear companies on trial.

87%

average shoe match across 41,000 beta takes

72%

of new workspaces made a usable image in their first 10 minutes

99%

lower cost per on-model image than a shoot day

86

System Usability Scale score in the final round

Before OnFoot

  • New product to on-model image 3–4 weeks
  • Catalogue with on-model images 6%
  • Cost per on-model image $80–150
  • Product accuracy check None
  • Generations per usable image 11
  • AI spend per usable image $1.75

With OnFoot

  • New product to on-model image Under 6 min
  • Catalogue with on-model images 80%
  • Cost per on-model image About $0.003
  • Product accuracy check Every take
  • Generations per usable image 1.3
  • AI spend per usable image $0.003

Trial funnel, beta cohort

Signed up100%
Uploaded a product84%
Generated takes79%
Exported an image64%
Converted to paid18%

Add-to-cart rate, design-partner listings

Weekly, eight weeks either side of launch. Same products, same prices.

8%6%4%2%0% OnFoot launch 6.8% with on-model images 2.2% packshot only

OnFoot turned on-model photography from a quarterly project into something a team does the day a product is ready.

The guided brief made creation easier for designers, while the internal shoe-match check helped them catch altered product details before approval. Together they made catalogue production more repeatable, without asking customers to evaluate AI accuracy.

Measuring success

One north star, and the signals underneath it

"Images generated" would reward bad takes and endless retries. The metric that matches the value a brand gets is a take good enough to sell with. Everything else is an input to it, or a guardrail on it.

North star metric

Listing-ready takes exported per active workspace, per week

Listing-ready means a shoe match of 7.5 or above, exported to a channel. It only goes up when the product does its job.

34per workspace per week in week 8 of beta, from 9 in week 1
Input metrics
ActivationTrial exports a take in 24 h64%
Pass rateTakes at 7.5 or above84%
Export ratePassing takes exported72%
Weekly useWorkspaces active in week 471%
GuardrailsFalse pass under 5%Cost per exported take under $0.25P95 generation under 60 s
HEART
Signal
Tracked event
Beta
Happiness
Satisfaction right after an export
One-tap rating after export
4.5 / 5
Engagement
Takes reviewed per session
Takes viewed, Loupe opened
6.2
Adoption
Workspaces using presets or refine
Preset used, take refined
58%
Retention
Workspaces still generating in week 4
Briefs submitted, by week
71%
Task success
Runs ending in an export, and time to it
Takes exported, by channel and score
72% · 4 min

Events were defined with the engineer during handoff, so the metrics existed on day one of beta instead of being reconstructed later.

From problem to outcome

Every feature traces back to a problem

I didn't add features just because I could. Each part of the product answers a specific problem I found, with a measure of whether it worked.

ProblemEvidenceDesign decisionOutcome
AI changes the product68% of AI outputs hallucinated shoe details; best model still 1 in 3Multi-angle references, locked prompt, automatic match check87% average match; 5% false passes
Teams can't review at scaleHuman review accuracy fell to 61% after 15 imagesScore bar on every thumbnail, issue note, Compare and Loupe"Spot the changed shoe" task: 2/5 → 4/4
Prompting is a skill barrier11 rewrites per usable image; 8 of 8 teams quitBrief panel, readable brief sentence, presetsFirst usable image: 9 min 40 s → 3 min 50 s
Images break across channels54% lost the shoe when cropped for a second channelRatio chosen before generating; shoe-weighted export cropsZero cropped-shoe exports in round 3
Fixes land on the wrong image3 of 5 testers refined the wrong takeRefine bar names its target; result saved as a new takeFix-and-export task: 2/5 → 4/4
Agencies mix client assetsStudio leads kept separate folders and logins per clientWorkspaces with roles, plans and separate history3 agencies in beta, 21 client workspaces

Honest imagery

An AI image that shows the wrong product is a misrepresentation, whatever it looks like.

What I didThe match check, plus an "AI-generated model" label in export metadata.

Representation

Defaults decide who appears on thousands of listings.

What I didGender and age are explicit choices with no hidden default; skin tone and body type are on the roadmap.

Photographers' work

Studio leads worried the tool would replace them.

What I didAgency workspaces and roles, so studios sell OnFoot output as a service.

Product confidentiality

Unreleased designs are commercially sensitive.

What I didPer-workspace storage; product photos never used to train shared models.

Auto-rejecting low scores

Version 1 hid takes below 7. Users lost good images the judge misread and stopped trusting the tool.

Lesson: let the system advise, let people decide.

AI-written briefs

I tried letting the AI expand short notes into full briefs. Results drifted from intent and nobody could tell why.

Lesson: keep what the user asked for visible and literal.

One take per run

Early runs made one image at a time. People retried endlessly and costs rose with no extra quality.

Lesson: give variety up front, priced openly.

What didn't work. Each failure shaped a principle that's still in the product.

Learnings & next steps

What I'd keep, and what I'd change

What I learned about designing with AI. The biggest decision was refusing to make OnFoot a better prompt box. A simple brief and a visible check did more for adoption than a better model.

Fewer choices make AI usable

Taking prompt writing away raised image quality and people's confidence at the same time. A few clear choices beat a blank text box, and gave the same look every run.

Let the AI say when it's unsure

The match score was a small build and the most-used feature. People trusted the tool more once it could say "this one's off, and here's why", not less.

Test the AI on real products

Every session used participants' own products. Stock or sample images would have hidden AI hallucinations affecting the shoe until launch.

What I'd do differently

Flag weak takes, never hide them

Version 1 hid takes scoring below 7. The check wrongly flagged 9% of good takes, so people lost images they would have used and stopped trusting the tool. I'd start with flagging from day one.

Name the score before testing it

"Quality 8/10" was read as a grade for the photo, not a check on the shoe. Renaming it "Shoe match" fixed it, but two rounds of results were muddied by the wrong word.

Test the AI with worse photos

Design partners' catalogues were beautifully shot, but many teams upload quick phone packshots. The match check struggled with those in beta. I'd add messy photos to the test set from week one.

← Back to Work