← Back to work Next Up:Product Image Platform

Outfit Generation Agent

Visit foxs.com ↗

AgentsVision ModelsImage GenerationShopifytrigger.devMeta Ads

Concepts

5,000+ outfit concepts per year

Time saved

~2,500 human hours per year

Ad volume

10x more Meta Ads — more shots at engaging content

Shop-the-looks are one of Fox's primary marketing units, running on foxs.com and in Meta Ads. Each look pairs a new arrival with a complete outfit so customers can buy the full set in one place.

Problem

Each shop-the-look takes meaningful merchandiser time to assemble — picking the anchor, deciding on a concept, and hand-picking compatible items category by category.
With ~1,500 in-stock SKUs and dozens of new arrivals weekly, the decision space for any single outfit is enormous, and the team can only ever ship a fraction of viable looks.

Solution

An agentic workflow running on trigger.dev anchors on each newly published Shopify item, iteratively builds a full outfit from the recent in-stock catalog, renders it as imagery, and routes it to the e-commerce team.

End-to-end pipeline

Daily Shopify Scan

Scans foxs.com daily for newly published items. Each becomes the seed for an outfit concept.

Iterative Outfit Build

Per category: an LLM text pass narrows to 16, then a vision pass over a grid picks the winner.

Image Generation

Selected products go to gpt-image-2 — a flatlay and an on-figure model shot per look.

Review + Publish

Each outfit posts to Microsoft Teams. The team edits in the dashboard and pushes to Meta Ads.

Inside the iterative outfit build

Category Candidates

Every in-stock SKU for the next slot — say ~300 dresses or ~200 shoes.

Text-Pass Shortlist

An LLM scores candidates on metadata against the concept and prior picks. Top 16 advance.

Single-Call Vision Grid

The 16 finalists composite into one 4×4 grid. One vision call picks the winner — not sixteen.

Append + Recurse

The winner joins the outfit and feeds the next category. Loops until every slot is filled.

The agent in flight

Per-category progress streams in live: text shortlist, vision grid analysis, selected product. Once every category is filled, gpt-image-2 renders the flatlay and on-model shot.

Generated looks ready for Meta Ads

Each completed outfit lands in Saved Ads with its product strip, flatlay, and on-figure shot. The team reviews, edits, and pushes to Microsoft Teams or Meta Ads from here.