Encoding stylist judgment into AI-generated fashion ads
Problem
Professionally produced editorial ads converted well but were a bottleneck to produce at scale. The company wanted to generate ads algorithmically, but needed a human QA layer — and expert judgment — to ensure the assets were high-quality and drove sales.
Solution
Labelbox Annotate was the core infrastructure for producing the signal and collaborating, yielding higher-quality signal and fast review of text and visual assets with the company's domain experts.
Result
Producing a set of ad assets used to take 2 weeks; it now takes three days, letting the company scale algorithmic asset generation and improve its core digital marketing.
A fashion ecommerce company generates personalized ads with AI. Labelbox captured stylists' judgment as the signal that trains the models, cutting asset production from two weeks to days.
The challenge
A leading fashion ecommerce company uses AI recommendation systems and wanted to improve its algorithmically generated ads. Creative assets drive conversion, adoption, and revenue, and one of its earliest data science teams worked closely with the marketing creative team on personalized ads. The company had millions of unstructured images to draw from. Professionally produced outfit matches converted well but were a bottleneck to produce at scale. The team wanted to generate those creative assets algorithmically and pair them with the right copy — which meant teaching models the taste and judgment that stylists bring.
The approach
The company used NLP and computer vision to generate both the copy and the outfit-combination images, with Labelbox as the core infrastructure for producing the signal and a human QA layer over every text and visual asset. Inside Labelbox, the team captured expert judgment by quizzing stylists on questions like “Would this layout be acceptable?” and “Would this outfit combination be acceptable?” — turning their recommendations into signal that trained the models to predict better. Copywriters used Labelbox as one centralized place for their work instead of scattered spreadsheets.
The outcome
A before/after comparison showed the payoff: producing a set of ad assets used to take two weeks and now takes a few days. A central platform to automate data import/export, speed human QA, and manage many users let the company scale its algorithmic approach. Today it runs nearly a thousand users across multiple teams and roughly 90 projects, producing more effective campaigns from one place to test and productionize AI.
Where this goes
Taste is domain expertise. Capturing a stylist's judgment as structured signal is how a model learns to make creative calls — the same pattern behind any specialist model trained on expert decisions.