A generated outfit looks like a shopping page from 2013: four or five product cut-outs on the same white seamless, trouser above shoe, bag hovering at the edge of the frame. A model posted to arXiv on 14 August by researchers at Hong Kong’s Laboratory for Artificial Intelligence in Design composes one of those grids in 0.08 seconds, drawing from a library of 251,008 catalogue items. Scarcity has left styling, and what it left behind is harder: the sentence explaining why this grid and not the ninety-nine after it. The engines solved supply. Nobody wrote the argument.

The paper reports 0.08 seconds per outfit at its cheapest search setting, 0.5 seconds at the configuration its authors call the optimal trade-off for interactive deployment, and 4.7 seconds at its heaviest. Eight copies of the cheap setting running side by side put a hundred finished looks a second in front of a single shopper. Alibaba showed the ceiling was reachable in 2019: POG, deployed on the Dida platform behind Taobao’s iFashion, generated about six million personalised outfits a day and had recommended them to more than 5.4 million users. POG’s authors note that Taobao’s own fashion experts were assembling thousands of outfits a day by hand, each reviewed before it went online.

Nothing in that pipeline produces a reason.

The August paper measures what the missing reason costs, though it does not frame the result that way. Thirty-two evaluators worked through 30 blind cases and made 928 choices between outfits from the new model, a vision-language model, and OutfitTransformer. The new model took 41.81% of those choices, the vision-language model 32.65%, OutfitTransformer 25.54%. Chance across three options is 33.3%, which puts the state of the art eight and a half points above indifference.

The circuit is closed, and worth tracing. The August model is tested zero-shot on iFashion, Alibaba’s own corpus, while Taobao has handed merchants an outfit-pairing tool that works inside the same property. Human stylists seeded the data, the data trained the engines, and the engines now fill the shelf the stylists used to fill.

The taste being modelled has an address, and it is a closed one. Polyvore Outfits, the benchmark everyone trains against, is built from sets that Polyvore’s users made by hand, before Ssense bought the site from Oath and shut it down the same day it announced the acquisition on 5 April 2018, leaving users six weeks to pull their own data before the sets disappeared for good. Every composed look in this literature descends from a community that argued about proportion in comment threads and no longer has a site to argue on. That inheritance is also why the outfits read as competent and interchangeable.

Volume of suggestion is not a new condition. Carmel Snow hired Diana Vreeland at Harper’s Bazaar in 1936 to write “Why Don’t You…?”, a column of propositions extravagant enough that S.J. Perelman parodied it in The New Yorker and enraged Snow in the process. One instalment proposed washing a blonde child’s hair in dead champagne, as they do in France. The ideas were unlimited and frequently absurd, and they landed because each arrived attached to a voice prepared to defend it.

Every fix the field has shipped since is supply-side. Compatibility scoring, aesthetic proxies, catalogue metadata, render fidelity: each assumes a shopper accepts the look once the look is good enough. Myntra’s My Stylist, drawing on roughly 450,000 styles, was reported at its 2023 launch to carry an engagement rate above 65%, per fibre2fashion, and that figure counts taps rather than convictions.

The argument a shopper is owed runs to one sentence. The trouser breaks clean at the ankle instead of pooling, the jacket carries enough shoulder to hold the volume beneath it, the shoe keeps the proportion honest. One line of that beneath a generated grid outperforms the ninety-nine grids stacked behind it. Composition costs nothing now, and the caption is the bottleneck.