AI Can Show You The Dress. But Can It Tell You If It Fits?
Updated: 5 hours ago
The $47 Billion Sizing Trap: Why Fashion’s Production Model Never Evolved
Apparel e-commerce has a $47.1 billion blind spot. According to a Coresight Research report sponsored by sizing technology firm Alvanon, U.S. online apparel and footwear returns reached an estimated $47.1 billion in 2025—representing a 23.4% return rate. In a survey of consumers who returned apparel purchased online, 69.8% cited size or fit as the primary reason.
In 1939, the U.S. Department of Agriculture began measuring American women to standardize clothing sizes. Its 1941 report, based on nearly 15,000 women, helped inform the federal sizing standard of 1958, though measurements from nonwhite participants were excluded and older women were sparsely represented. More than six decades later, the 2002–2003 SizeUSA survey used 3D body scanners to measure roughly 10,000 Americans, revealing significant changes in body proportions.
From tape measures to 3D scanners to AI, the technology for measuring us has advanced dramatically. Yet most clothing is still developed around reference bodies: dress forms and live fit models whose approved measurements shape patterns and grading.
Eighty-five years after the USDA’s report, AI can create remarkably detailed digital representations of our bodies. But has the industry changed how it makes clothes—or simply become better at matching customers to the sizes it already produces?
The High Cost of the Virtual Try-On Illusion
Artificial intelligence has made shopping for fashion remarkably easy. Shoppers can ask ChatGPT, Gemini or a growing roster of generative retail tools to curate silhouettes, occasions and price points in seconds.
With virtual try-on (VTO) technology, customers can now see themselves wearing garments before checking out—or so the marketing promises.
A virtual try-on tool can place a garment onto a shopper’s photograph, creating a convincing image. The dress falls beautifully. The proportions appear flattering.
Yet what appears on a screen frequently bears little resemblance to physical reality. Even as an apparel veteran with three decades in product development, I still find myself ordering multiple sizes of the same item to find one that works, directly contributing to retail’s margin-eroding return cycle.
A generated image may suggest a silhouette without accurately accounting for fabric stretch, armhole depth, waist placement, pant rise, or how a garment behaves in motion.
That does not diminish the commercial value of AI visualization. As I reported previously, Zelig’s AI-powered styling tools helped REVOLVE increase conversion and average order value while reducing returns. These technologies can improve product discovery and purchasing confidence.
But visualizing an outfit and predicting physical fit are different challenges.
The question is no longer whether AI can improve shopping. It already has. The challenge is whether it can help shoppers determine how clothing will actually fit before they buy it.
Vestir, which collaborated with ETH Zurich on virtual try-on research, combines 3D reconstruction, physics-based fabric simulation and generative AI. Founder Joanne Filbin says creating a digital avatar is relatively straightforward; accurately digitizing the garment remains “a massive challenge.”
“Gen AI can make something look real. But what these Gen AI solutions don’t have is the ability to solve the size and the fit,” Filbin told me.
Her research collaboration achieved reported avatar accuracy within 2 millimeters and scanned-garment accuracy within 1 centimeter using an eight-camera setup. However, those results do not establish the accuracy of Vestir’s newer consumer photo-to-3D process, where uncalibrated smartphone cameras, uncontrolled lighting, and inconsistent user angles introduce substantial noise.
The challenge is translating convincing digital images into reliable predictions of physical fit.
The Zalando Test: When AI Closes a Zipper Reality Can’t
Zalando offers an instructive example of how far virtual fitting rooms have advanced. In a 2023 pilot involving 22 jean styles, shoppers created 3D avatars and viewed heatmaps indicating where different sizes might fit tightly or loosely. In subsequent virtual fitting room tests, the company reported return-rate reductions of up to 40%. Its broader size-and-fit tools reportedly prevented 8% of size-related returns in 2025. These results cover different initiatives and are not directly comparable.
Unlike third-party sizing providers that rely on garment information provided by participating brands, Zalando operates its own garment digitization facility. According to a Modaes report, its Berlin Fitting Lab, led by Head of Size & Fit Niclas Grabau, uses a modified industrial scanner equipped with 120–140 cameras. The equipment captures challenging areas of the garment, particularly the crotch, and produces digital files annotated with garment components, including waistbands, pockets, fronts, and backs.
I decided to test Zalando’s virtual fitting room myself. After creating a digital body profile
using my smartphone, I compared its measurements with those I had taken manually. Several were close, differing by just 1–2 centimeters (0.4–0.8 inches). The thigh measurement was 4 centimeters (1.6 inches) larger, and no waist measurement was provided.
I then virtually tried on Levi’s jeans. Zalando recommended a size 26×27, consistent with the size 26 I wear in Levi’s. The avatar showed some looseness around the hips and thighs, much like my actual jeans. When I selected larger sizes, including 32×27, the technology convincingly illustrated excess width and length.
But the real test came when I selected a size 23×30.
The jeans appeared tighter around the waist and hips, yet the avatar showed the waistband and fly fully fastened. The product specifications listed 85% cotton, 14% polyester, and 1% elastane. I know from experience that I could not squeeze into a size 23, much less zip it up. Yet the virtual fitting room showed me wearing it.
Zalando’s technology is impressive, particularly in visualizing differences between sizes. But my test revealed an important limitation: showing how a garment looks on a digital body is not the same as establishing whether the physical garment can actually be worn.
The Measurement Dilemma: Anatomy vs. Algorithmic Assumptions
I also tested TrueToForm, a sizing-recommendation platform, and Size Stream’s Formcut, a made-to-measure measurement app. Both use smartphone-based body scanning.
Having spent more than 30 years working across apparel product development, sourcing, technical design, and production, I measured myself manually with a tape measure, as I would for a fit session. I wore the same clothing for all three scans, including Zalando’s. Some measurements differed from mine by one to three inches.
These were exploratory tests involving one body, not a controlled comparison or ranking. The systems also serve different purposes: Zalando digitizes garments within its own operation, while TrueToForm and Size Stream work with brands that provide varying levels of garment data.
Were the scans inaccurate, or were the systems and I measuring different points on the body?
TrueToForm co-founder Janice Tam explained that posture, breathing, tape compression, and measurement definitions can all influence results. The natural waist, for example, may differ from the location where a shopper measures the waist of her trousers.
Jay Rudman, CEO of Size Stream, made a similar point: manual tape measurements also vary with placement, compression, and the person taking them. Size Stream validates its technology against high-precision booth scanners rather than treating manual measurements as an unquestioned reference.
In my Size Stream - Formcut test, the waist and hip measurements raised particular questions because they differed from my tape measurements. But those discrepancies alone do not establish whether the scan or my manual measurements were closer to the relevant anatomical standard.
Both systems also demonstrate capabilities beyond generating numbers. TrueToForm translates body-shape information into brand-specific recommendations, while Formcut captures measurements intended to support made-to-measure garment development.
In my TrueToForm shopping tests, the system recommended a Small at Rainbow Shops and a size 6 at New York Dress, both larger than my usual size selection. Different brands use different size specifications, but the recommendations raised questions about how garment information and inventory availability affect the result.
Tam confirmed that TrueToForm considers the closest available size and can indicate where clothing may fit more loosely or snugly. The best available size, however, is not necessarily the best-fitting size—a crucial distinction for retailers who mistake inventory matching for genuine fit satisfaction.
She said the system combines body-shape information with brand size charts, garment specifications, and purchasing data to improve recommendations.
“The system performs best where we have the most robust scan data to train on,” Tam explained.
Her larger ambition goes beyond helping consumers select from existing sizes: “The goal is not just a sizing widget for shoppers; the vision is bigger than that. Like we’re trying to improve the way that clothing is made in order to change the way that clothing fits.”
Bridging Code and Cut: What the Factory Floor Knows
“The scanning technology is actually probably the easiest part of the entire equation,” Rudman told me.
His point is that body measurements are only the beginning. A factory still needs to translate those measurements into patterns, garment specifications, construction allowances, and intended fit.
Size Stream is owned by TAL Apparel, a global apparel manufacturer. That relationship connects body-scanning technology with the practical requirements of garment production.
“Body measurements and garment measurements are totally different, and they depend on the brand and the tech pack and ... the fabric and everything else,” Rudman explained.
He calls the translation sizing logic. A fitted jacket cannot simply reproduce the customer’s chest measurement. It requires ease, fabric allowances, and decisions about the intended silhouette.
Size Stream’s made-to-measure applications combine scanning, garment configuration, fit preferences and manufacturing specifications. Rudman said the company typically sees remake rates of approximately 5%, a company-reported figure for made-to-measure operations rather than a ready-to-wear return rate.
The bigger opportunity is whether the same information can improve standard sizes for thousands of customers.
From Scans to Sourcing: DXL and the Upstream Data Pipeline
Destination XL Group (DXL), the men’s big-and-tall specialty retailer, has worked with Size Stream since 2019 on FiTMAP, a body-scanning system designed to help customers navigate sizing differences across brands.
FiTMAP captures 243 body measurements and provides personalized recommendations. According to DXL’s fiscal 2025 annual report, the company covered 29 brands and completed more than 63,000 customer scans. By September 2026, DXL reported more than 150,000 customer engagements, a broader metric than completed scans.
DXL has reported higher conversion, average order value, and purchase frequency, along with lower return rates among FiTMAP users. The company has not publicly established how much of those differences were caused by the technology.
The more consequential opportunity may lie upstream: could the accumulated measurements of real customers help brands refine fit blocks, grading rules, assortment planning, and production specifications? Done right, that feedback loop wouldn’t just improve how clothes fit—it would dismantle the root causes of apparel’s return rates and costly overproduction.
Size Stream’s published DXL case study highlights the potential value of body-profile data for inventory decisions, although DXL has not publicly confirmed that it has redesigned its production fit blocks using FiTMAP data.
Another example comes from Bershka, part of Inditex. In the Coresight report sponsored by Alvanon, the brand was described as using digital avatars and corresponding physical mannequins to align sizing standards across development teams and suppliers, with a reported 10% reduction in returns. That result is presented in sponsored research and should be treated as a company-associated case study, not independently verified causation.
The opportunity is compelling: rather than merely recommending the closest available size, retailers could use real customer body data to improve the garments they manufacture.
The Missing Half: Digitizing Fabric Mechanics and Ease
Even accurate body measurements cannot predict fit without equally reliable information about the garment.
A tech pack may specify finished measurements, construction, fabric composition, and grading rules. But those specifications do not necessarily reveal whether a waistband will dig into the stomach, a jacket will pull across the shoulders, or a dress will ride up when someone sits.
Audrey-Laure Bergenthal, founder of French fashion technology company Euveka, argues that the industry has focused too heavily on the body while overlooking the garment.
“Scans and measurements already give a reliable body model. But the body is the easy half: an avatar says nothing about fit until you know the garment as precisely as you know the person,” she told me. “That missing half is what we provide: the full garment fit intelligence.”
“Points of measure (POMs) and grading describe a garment flat on a table, not on a body,” Bergenthal explained. “Fit is the relationship between the two: ease, tension, fabric behavior, and the fit the brand intended.”
Euveka’s Emili robotic mannequin adjusts its dimensions to represent different body morphologies, including variations in breast volume and cup size, allowing brands to evaluate physical garments against a wider range of body shapes than a conventional fixed dress form.
One example came during Beyoncé’s 2023 Renaissance World Tour. According to Bergenthal, Euveka reproduced the performer’s individual body measurements on its adjustable mannequin to support the development and fitting of a custom Mugler bee costume.
The Beyoncé example illustrates how physical fit can be evaluated against an individual’s measurements rather than a standardized dress form. Euveka’s applications in luxury and ready-to-wear product development extend that principle across different body morphologies. The significance is not simply the ability to reproduce a body shape, but to observe how an actual garment fits that shape—information a digital body scan alone cannot provide.
Testing for Real Bodies: Beyond the Static Fit Model
Imagine a shopper’s at-home 3D scan connecting directly to a garment’s physical fit profile before purchase.
This does not require custom mannequins or endless micro-sizes. Instead, brands test physical garments on adjustable robotic forms across representative body archetypes, mapping real fabric stretch, tension, and ease. When customers scan themselves at home, algorithms cross-reference their personal proportions against that garment data—predicting exactly how the piece will drape and move.
Bergenthal envisions a future Morpho-Passport, a universal reference that works across brands: “Tomorrow you won’t need to know your size. You will simply have your reference, the way a color has its Pantone”. That points toward the ultimate industry shift: connecting home body data with garment physics upstream, rather than leaving fit to chance at checkout.
The Executive Verdict: The Sizing Widget Is Not Enough
Not reliably enough—yet.
In my tests, Zalando convincingly showed me oversized jeans but also displayed a fully fastened size 23 that I know I could not wear. Other scanning systems generated useful body information while producing measurements that differed from my tape by inches.
These were not controlled accuracy studies. But they revealed the distance between a persuasive digital experience and dependable physical fit.
AI has made extraordinary progress in understanding and visualizing the body. The harder challenge is connecting that knowledge with how garments are constructed, how fabrics behave, and how sizes are developed.
Eighty-five years after the USDA began standardizing women’s measurements, fashion has far more sophisticated tools to understand the human body.
The real opportunity is not simply to recommend which size a customer should buy. It is to use what we learn about real bodies to make better-fitting clothes in the first place—and, finally, to dismantle retail’s $47 billion return problem at its root.
Comments