Does virtual try-on show how a garment will fit?

Not yet. A 2026 SIGGRAPH paper from the University of Washington and Google Research found that existing try-on methods show a well-fitted garment regardless of the garment or body size, and fit-aware try-on is still a research frontier built on synthetic data. Treat image try-on as styling and use size advice for fit.

Has AI been shown to reduce fashion returns?

Yes, by size advice rather than images. In a Zalando and TU Dortmund A/B test with more than 132,000 customers per group, article-level size advice cut the size-related return rate by 4.5% relative. We found no peer-reviewed study showing image try-on reduces returns.

Does AI make designers more creative?

It helps individuals but narrows the group. A 2024 Science Advances study found AI ideas made individual work more novel and useful but made outputs more similar to each other. Fashion practitioner studies report efficiency gains alongside worries about dominant aesthetics.

Are AI fashion images biased?

Research says yes unless steered. A 2026 audit of two commercial image models found over 96% of people generated from neutral prompts were white-presenting, and a NAACL 2025 paper documented weight bias in DALL-E 3 images.

Does AI reduce overproduction in fashion?

It has not been shown. We found no peer-reviewed study measuring a net reduction, while image generation uses over 60 times the energy of text generation per a 2024 ACM FAccT study.

Fashion AI research: what universities say

Academic research on fashion AI tells a more sober story than the launch announcements. Virtual try-on images now look photoreal and cost far less to make, but researchers at the University of Washington and Google wrote in 2026 that today's try-on tools show a "well-fitted" garment whatever the size of the body or the garment. The returns reductions that have been measured in controlled tests came from size advice built on purchase and return data, and they are in single digits. Randomised field experiments in online retail found sales effects from nothing to 16%, depending on the job AI was given. Studies of creative work find that AI lifts individual output but makes work more alike, and image models default to a narrow look unless steered.

This is part two of our fashion AI research series and the first of five deep dives. It covers what universities and research labs have published since 2024, with older papers where they remain the basis, and separates peer-reviewed results from preprints, lab benchmarks and opinion.

How good is virtual try-on, according to the research?

The lab progress is real. Google and the University of Washington's TryOnDiffusion (CVPR 2023) showed a single diffusion network could warp a garment onto a person while keeping its detail through large changes in pose and shape. KAIST's IDM-VTON (ECCV 2024) improved how faithfully the garment's own identity survives. By 2025, CatVTON (ICLR 2025) trained only about 50 million of its 899 million parameters and used over 49% less memory than other diffusion try-on methods, without needing pose estimation at inference. Try-on has become cheap enough to deploy widely.

Three lines of research say what that progress does not mean.

It does not show fit. The FIT dataset paper from the University of Washington and Google Research (SIGGRAPH 2026) states that existing try-on methods "default to generating well-fitted results regardless of the garment or person size", because no dataset recorded precise body and garment measurements, least of all ill-fitting ones. Its answer is more than a million synthetic training examples across 168 body shapes from XS to 3XL. FitVTON (June 2026) makes the same point: diffusion try-on treats the task like filling in a 2D picture, "prioritizing texture preservation over physical plausibility". SV-VTON (2025) adds that methods overlook garment size variations altogether. Fit-aware try-on is a 2026 research frontier, not a shipped capability.

The benchmark scores do not match what people see. VTBench (2025) found that "current metrics inadequately reflect human perception", that most test sets are shot indoors, and that results drop in real-world scenes. A NeurIPS 2025 paper from USTC and HiDream.ai, VTON-VLLM, found try-on models that score well on the standard image metrics still "often fall short" of users' expectations on texture and on how the garment sits on the body. A survey of image-based try-on had already named "a gap between current research progress and commercial applications".

Fine detail is the weak spot. Leffa (2024) starts from the observation that generators distort logos, text and patterns. A 2025 comparison of large language models on garment images found they struggled with buttons, necklines and exact colours. For jewellery, which is almost all fine detail, that matters; we found no academic benchmark for jewellery try-on at all.

What has been measured about size, fit and returns?

The strongest commercial evidence in the academic literature does not come from images. It comes from size advice built on what customers bought, kept and returned.

Study	Design	Result
Zalando and TU Dortmund, MultiFlags (IC3K 2025)	A/B test, over 132,000 customers per group, 12 countries	Size-related return rate down 4.5% relative; 7.3% for full-body garments
Amazon, Recommending Product Sizes (RecSys 2017)	Offline evaluation, then live A/B test	17 to 21% better offline; +0.49% fit transactions live
Fang, Sarvary and co-authors, field experiments in online retail (2025, revised 2026)	Randomised experiments with millions of users, seven workflows	Sales effects from none to +16.3% by workflow; return rates and ratings did not worsen

Two lessons follow. The Amazon paper is an honest early example of large offline gains shrinking in production, which is why a pilot must be judged on live results. And the Zalando paper notes there is no ground truth for anyone's "true size": article-level advice works alongside personal recommendations rather than replacing them. Zalando's own corporate figures, outside the paper, put the avoided share of size-related returns at more than 8% in 2025.

The field experiments by Fang and colleagues are the best causal evidence on generative AI in commerce we found. Gains came from conversion rather than bigger baskets, and were larger for less experienced shoppers. The effect depended on how much AI added to what the platform already did. That is a useful template for any brand's pilot.

What does research say about AI and creative work?

The findings on creativity are consistent: AI helps individuals, and narrows the group.

Homogenisation. In a pre-registered study published in Science Advances (2024), Doshi and Hauser found that writers given AI ideas produced stories rated 8.1% more novel and 9.0% more useful, but the stories were more similar to each other. They call it a social dilemma. For a design team, the parallel is a collection that is individually polished and collectively generic.
The jagged frontier. In the Harvard, Wharton, MIT Sloan and BCG experiment with 758 consultants, AI users finished 12.2% more tasks, 25.1% faster, at over 40% higher quality on tasks inside AI's capability, and were 19 percentage points less likely to be right on a task outside it. Merchandising and design work has both kinds of task.
Practitioners. A Hong Kong Polytechnic University study, evaluated by 30 fashion practitioners and two companies, found a generative design system sped up design and communication but needed good data and raised copyright questions. A 2026 survey of 93 fashion professionals in AI and Ethics found AI valued for efficiency, but seen to reinforce dominant aesthetics and push aside non-Western and niche ones. Researchers at DTU and Padua warn that over-reliance "may reduce authenticity", especially in luxury.

London College of Fashion's Fashion Innovation Agency has worked with AI since 2019; its director Matthew Drinkwater describes the aim as AI that will "support and enhance creativity, not replace it", and calls AI "a mirror" of the data it is given.

Is bias in AI fashion imagery a real problem?

Yes, and it is documented. A 2026 preprint testing 3,200 photorealistic images from neutral prompts in two current commercial image models found over 96% depicted white-presenting people. A Findings of NAACL 2025 paper analysed 4,000 DALL-E 3 images for anti-fat and pro-thin bias, noting no earlier study had looked at weight bias in AI images. Fashion academics connect this directly to the industry's history: London College of Fashion's 2026 International Symposium on AI in Fashion included papers on racially skewed outputs from neutral prompts and on "the standard body" reproduced by AI.

For a brand, the practical point is simple: generated models and lookbooks will default to a narrow look unless you specify otherwise and review what comes out.

Does AI make fashion more sustainable?

The research does not yet show it. AI is often pitched as cutting overproduction and sampling, but we found no peer-reviewed study that measures a net reduction in fashion. On the other side of the ledger, Luccioni, Jernite and Strubell measured image generation at an average of 2.907 kWh per 1,000 images, over 60 times the energy of text generation (ACM FAccT 2024). The 2026 LCF symposium programme included a paper titled "Built to Overconsume: AI and Overproduction in Indian Fashion". Sustainability claims for AI should be specific and should count the compute.

Where is academic attention moving?

The agenda has shifted from tools to consequences. The 8th Computer Vision for Fashion, Art and Design workshop at ICCV 2025, with organisers from universities and Zalando, was subtitled "Bridging Creativity and Responsible AI". LCF's symposium tracks covered overproduction, disability and bias, labour in AI supply chains, race in AI imagery and craft knowledge, including a contribution from NIFT India. A 2025 LCF study on teaching AI in fashion management found students more confident with AI, with ethical concerns and unclear rules still unresolved.

The wider business data from universities points the same way. Stanford's AI Index 2025 found organisational AI use rose to 78% in 2024, but most companies reporting revenue gains in marketing and sales put them under 5%; its 2026 edition puts use at 88%. MIT's Project NANDA report that about 95% of generative AI pilots showed no P&L impact is widely quoted; it is preliminary, not peer reviewed and reported with differing sample sizes, so treat it as a warning rather than a measurement.

How strong is this evidence?

Claim	Evidence type	Confidence
Image try-on does not show real fit	Peer-reviewed (SIGGRAPH 2026) and preprints	High
Standard try-on metrics diverge from human judgement	Peer-reviewed (NeurIPS 2025) and benchmark preprints	High
Size advice cuts size-related returns by single digits	Live A/B test at scale (peer-reviewed proceedings)	High for that setting
Generative AI lifts online sales by 0 to 16% depending on workflow	Randomised field experiments (preprint)	Medium to high
AI homogenises creative output	Peer-reviewed experiment (writing, not fashion)	Medium for fashion
Image models default to narrow looks	Peer-reviewed and preprint audits	High
AI reduces overproduction	No peer-reviewed measurement found	Unproven

What should a fashion or jewellery brand take from the research?

Do not promise fit from a try-on image. Sell try-on as styling and visualisation. For size, use advice built on garment size charts and your own purchase and return data; that is where measured reductions in returns come from. We explain the approach in the right size, the first time.
Judge try-on with people, not metrics. Check logos, engraving, stone settings and metal finish with your own team before anything reaches a shopper.
Expect single-digit gains and test for them properly. Run controlled tests, and track returns and ratings alongside conversion. Part nine sets out the design.
Steer imagery for representation. Specify body size, skin tone and age; review outputs; never let the default be the brand.
Use AI to widen the creative funnel, then curate. Let designers decide, and start from your own evidence rather than a blank prompt, as in from demand signals to new designs.
Make sustainability claims carefully. Count the compute, and claim only what you measure.

Some of these lines match choices Ateliq has made. Its camera never guesses a size: shoppers tap the sizes they wear, and the stylist uses them to check stock. It never infers body shape, race, ethnicity or weight. Its photo-real try-on is made from the brand's own product photos, and the photographer agent never changes the piece. FabFit, its size engine, is in early access and not yet generally available.

Next in the series: what investors are betting on, and what they admit they do not know.

Fashion AI in the lab: what universities and research labs say