AI in fashion covers three separate things: making the images, changing how people shop, and changing what people wear. Only the first has broad production evidence behind it. The second works but most published numbers come from the companies selling the technology. The third barely moved. Treating all three as one roadmap is how brands lose a year.
You have read the list before. Design, sustainability, supply chain, personalization, virtual runways, trend forecasting, chatbots. Every item true in isolation, none of them ranked, and no way to tell which one your brand should touch first.
This guide sorts the same category by a single question: who published the number. That test separates what is running in production from what is running in a press release, and it puts the three layers in an order you can act on.
Key Takeaways
35% of fashion executives already use generative AI for functions including image creation, copywriting, customer service and product discovery (BoF-McKinsey, The State of Fashion 2026). Creative production is the layer with public, audited evidence. Zalando reported roughly 70% of its Q4 2024 editorial campaign imagery was AI-generated, cutting production from six to eight weeks down to three to four days (Reuters, May 2025). Nearly every virtual try-on returns statistic is published by a try-on vendor. Treat the 15% to 48% reduction range as a marketing claim until an independent number exists. Apparel is the highest-returning ecommerce category at 20% to 40%, with Coresight Research putting online apparel at 24.4%, and fit is the largest single cause. Digital-only clothing did not become a category. Peer-reviewed work describes sharp declines in transaction volume, not growth. As of 2 August 2026, EU transparency obligations apply. The Commission’s 20 July 2026 Guidelines confirm a realistic image of a fictitious person counts as a deepfake, so AI fashion models need labelling even when nobody real is depicted. Start where the recurring spend already sits. For most fashion brands that is photography, not forecasting.
What does AI in fashion actually mean in 2026?
AI in fashion means three distinct transformations that get bundled into one phrase. The create layer generates product and campaign imagery from a garment photo. The shop layer changes discovery, sizing and recommendation inside the storefront. The wear layer covers digital-only clothing and virtual goods. Each sits at a different maturity level, and the evidence supporting each differs by an order of magnitude.
The bundling is the problem. A guide that puts demand forecasting, virtual runways and flat-lay generation in the same list implies they are comparable decisions. One needs years of clean sales data and a team of analysts. One needs a platform partnership. One needs a photo of a jumper and an afternoon.
Adoption data shows where brands actually landed. The State of Fashion 2026 reports 35% of executives using generative AI, concentrated in marketing and consumer-facing functions, with other teams lagging. The same report notes that most companies remain stuck in pilot mode.
That concentration is not an accident. It is the market discovering which layer pays.
Layer one: create, where the evidence is public
This is the only layer with audited numbers from operators rather than suppliers.
Zalando is the clearest case. In the fourth quarter of 2024, roughly 70% of the retailer’s editorial campaign images were AI-generated. Its VP of Content Solutions reported imagery production falling from six to eight weeks to three or four days, at about 90% lower cost (Reuters, May 2025). Zalando was explicit that the argument was speed and relevance rather than output quality, and that its creative team stayed busy.
Those are operator numbers, published by a listed company describing its own operation. That is a different class of evidence to a vendor case study.
What the create layer covers in practice:
| Job | Input you need | Evidence level |
|---|---|---|
| Packshots and flat lays | One garment photo | Strong, in production at scale |
| On-model imagery | A flat or mannequin shot | Strong, in production at scale |
| Campaign and ad variants | Existing brand assets | Strong, in production at scale |
| Fashion video from a still | A finished still | Moderate, growing fast |
| Print and pattern concepting | A design team already in place | Moderate, design-led |
The economics are the reason this layer moved first. A one-day on-model shoot runs several thousand dollars for a few dozen finished images, and a fashion brand with a monthly drop calendar pays that repeatedly. Photography is recurring spend with a fixed unit cost, which is exactly the shape that automation attacks well.
The practical constraint is that output quality depends on the input photo, not the prompt. Garment texture, drape and colour survive or fail based on what you shot. Our guide to creating fashion visuals with AI covers which input each garment type needs.
For teams choosing where to run this, DesignerBox Model Studio builds product stills, on-model shots and fashion video from the same source photo, and the product photography model comparison runs one brief through every available model so you can see the difference before committing.
Layer two: shop, where you should check the byline
The shop layer is real. Virtual try-on, size recommendation and AI-assisted discovery are shipping inside major storefronts. The problem is not whether they work. The problem is that almost every published figure about how well they work was published by a company selling them.
Search for the impact of virtual try-on on returns and you get a range from 15% to 48%. Trace each number to its source and you find try-on vendors, retail-tech blogs citing try-on vendors, and one study of over a million shoppers published by a digital fashion company that sells try-on. No independent, peer-reviewed measurement of virtual try-on’s effect on apparel returns surfaced in this review.
That does not mean the technology fails. It means you cannot plan a budget on a number whose author has a stake in it.
The underlying problem is well documented and worth acting on regardless. Apparel is the highest-returning ecommerce category. Coresight Research puts online apparel returns at 24.4% on average, with broader industry benchmarks placing the range at 20% to 40% and women’s fashion at the top end. Fit and sizing are consistently reported as the largest single driver, though published estimates of their share vary widely because surveys, retailer reason codes and category studies each measure something different.
Here is the distinction that matters when you evaluate a try-on pilot. Try-on answers a styling question: does this look right on a body like mine. Fit answers a dimensional question: will this actually fit my body. Those are different problems, and the bigger returns driver is the second one. A visualisation tool that does not measure you is a confidence tool, not a fit tool. We go deeper on that split in virtual try-on fit accuracy.
The honest way to run this layer is to instrument it yourself. Pilot on one category, hold a control group, measure your own return rate over a full season, and treat the vendor’s number as a hypothesis rather than a forecast. DesignerBox includes virtual try-on on Premium and above if you want to generate the on-model variants without a separate platform contract.
Layer three: wear, the one that did not arrive
Digital-only clothing, NFT wearables and metaverse fashion were the headline of nearly every AI fashion guide written in 2022 through early 2025. They remain the headline in guides that were never updated.
The category did not scale. Peer-reviewed work on digital fashion in the metaverse describes a speculative dimension that complicates positioning digital fashion as a stable consumer product, and periods of sharp decline in transaction volume following market downturns, which makes revenue projection unreliable for brands and platforms alike (Taylor and Francis, 2025). A Berkeley California Management Review assessment of luxury in the metaverse published in May 2025 framed the question as digital renaissance or fading mirage, which is not the framing a growing category attracts.
Market-forecast reports still project large numbers for metaverse fashion. Those come from commercial research mills, they are projections rather than measurements, and they should not carry the same weight as a company reporting what it actually did.
The useful takeaway is not that digital fashion is worthless. It is that a guide leading with virtual runways and digital-only garments is telling you about 2022. Check the publication date before you build a roadmap from it.
Why a 2025 AI fashion guide points you the wrong way
Three things changed between early 2025 and now, and none of them appear in a guide written before them.
Model generations turned over completely. Capability claims about resolution, garment fidelity and video length made eighteen months ago describe systems that have been replaced.
Adoption concentrated. The 2025-era guides spread attention evenly across the value chain. The 2026 data shows brands concentrating in marketing and consumer-facing creative, with the rest stuck in pilots.
The legal position changed on 2 August 2026. That one deserves its own section.
What changed on 2 August 2026
Article 50 of the EU AI Act became applicable on 2 August 2026. It requires that AI-generated synthetic image, audio, video and text output be marked in a machine-readable format and detectable as AI-generated, and it places a disclosure obligation on deployers, which includes the brands and agencies publishing the content, not only the model providers.
The clarification that matters most to fashion arrived on 20 July 2026, when the European Commission adopted the final Guidelines on Article 50 transparency obligations. Those Guidelines confirm that a realistic synthetic depiction of a fictitious but natural-looking person is a deepfake under Article 3(60), even where no identifiable rights-holder is involved. Content that looks like a real person must be labelled even when no real individual is depicted and no deception was intended.
Read that against what fashion brands actually generate. An AI fashion model is an invented person who looks real. Under the Guidelines that is a labelling case, not an exemption.
Two timing details are worth holding. The obligations apply from 2 August 2026. The AI Omnibus provisional agreement of May 2026 gives generative systems already on the market before that date until 2 December 2026 to meet the machine-readable marking requirement in Article 50(2).
The Commission’s Code of Practice on Transparency of AI-Generated Content is voluntary and functions as a compliance route rather than a rule. It states plainly that no single labelling technique works in all cases.
Practice is already ahead of enforcement in one respect. Where brands disclosed AI imagery only in small print, the reaction was worse than where the labelling was visible. Prominent disclosure is becoming the expectation rather than the minimum. We cover the operational side in labeling AI-generated fashion images.
How to decide what to run first
Apply the evidence test before the feature list.
Ask who published the number. An operator reporting its own results is stronger than a vendor reporting its customers’ results. A peer-reviewed study is stronger than both. A market-forecast projection is not evidence of anything.
Ask what the recurring spend is. Automation returns most where a fixed cost repeats. For a fashion brand with a monthly drop, that is photography. Forecasting saves money once you have years of clean data, which is a different project on a different timeline.
Ask who can start without permission. Visual production runs from a garment photo, by one person, this week. Try-on and personalization usually wait on your ecommerce platform’s roadmap. Forecasting waits on a data function.
Ask what you can measure yourself. Set your own baseline before the pilot. A control group and one full season beats any published percentage.
Ranked against those four tests, the order comes out the same for most brands: creative production first, storefront features when your platform ships them, forecasting only at enterprise scale, digital-only garments not at all yet.
If you want to start on the first one, DesignerBox puts 13 image and video models behind one subscription. Basic is $15 a month for 500 credits, the free plan includes 112 credits with no card, and the commercial licence starts at Pro. The fashion OOTD workflow reruns the same setup for each new drop rather than rebuilding it per campaign.
FAQ
What is AI used for in the fashion industry?
Mainly creative production. The State of Fashion 2026 reports 35% of fashion executives using generative AI, concentrated in image creation, copywriting, customer service and product discovery. Supply chain forecasting, design concepting and personalization all exist, but adoption there is thinner and more often stuck at pilot stage.
Do AI fashion models have to be labelled?
In the EU, yes, in most cases. Article 50 of the AI Act applies from 2 August 2026, and the Commission’s 20 July 2026 Guidelines confirm that a realistic image of a fictitious but natural-looking person counts as a deepfake under Article 3(60). The labelling duty applies even though no real person is depicted. Check your own jurisdiction and platform policies as well.
Does virtual try-on reduce returns?
Published figures suggest reductions between 15% and 48%, but nearly all of them come from companies that sell virtual try-on. No independent measurement surfaced in this review. Fit and sizing are the largest reported cause of apparel returns, and a visualisation tool that does not measure your body addresses styling rather than fit. Run your own pilot with a control group.
Is AI cheaper than a fashion photoshoot?
For repeat volume, generally yes. Zalando reported roughly 90% lower imagery cost and a drop from six to eight weeks to three or four days (Reuters, May 2025). The saving depends on your cadence. A brand shooting once a year sees little; a brand with a monthly drop sees the difference every month.
What happened to digital fashion and the metaverse?
The category did not scale. Peer-reviewed work describes sharp declines in transaction volume and difficulty positioning digital fashion as a stable consumer product. Large market-size projections still circulate, but they come from commercial forecast reports rather than measured revenue. Any guide leading with virtual runways is describing 2022.
Which AI fashion use case should a small brand start with?
Product and on-model imagery. It is the only layer that runs from an asset you already have, needs no platform partnership or data team, and attacks a cost you pay again every drop. Everything else either waits on your ecommerce provider or needs years of clean sales data.
Adoption, regulatory and returns figures verified from BoF-McKinsey, the European Commission, Reuters and Coresight Research as of August 2026. Virtual try-on performance figures are vendor-published and flagged as such. Individual results vary.
Sources
- BoF-McKinsey, The State of Fashion 2026, mckinsey.com, accessed August 2026
- European Commission, Guidelines on transparency obligations under Article 50, adopted 20 July 2026, digital-strategy.ec.europa.eu, accessed August 2026
- European Commission, Code of Practice on Transparency of AI-Generated Content, digital-strategy.ec.europa.eu, accessed August 2026
- EU Artificial Intelligence Act, Article 50 and Article 3(60), artificialintelligenceact.eu, accessed August 2026
- Reuters, Zalando uses AI to speed up marketing campaigns and cut costs, May 2025
- Coresight Research online apparel return rate, cited via industry benchmark reporting, 2026
- Examining Digital Fashion’s Functions and Performers in the Metaverse, Taylor and Francis, 2025
- Luxury in the Metaverse: A Digital Renaissance or a Fading Mirage, California Management Review, Berkeley, May 2025
