A virtual try-on image can show a look without proving the fit

Research on garment-image synthesis suggests a useful shopping preview, provided appearance, sizing and usage rights stay separate.

A virtual try-on image can show a look without proving the fit

A shopper sees their photograph wearing a striped jacket. The preview is convincing: the collar follows their shoulders, the sleeves bend naturally and the stripes remain recognisable. That image may help them explore a style. It does not, by itself, establish whether a particular size will close comfortably, how the lining feels or how the fabric behaves in motion.

For retailers considering image-based virtual try-on, the central product decision is what the preview is allowed to mean. This analysis examines research from 2022–2024, checked on 17 October 2026. It concerns image synthesis, not a tested sizing service. Its scope is international research, including the South Korean KAIST researchers behind IDM-VTON, rather than one national retail market. The papers and datasets are not proof of commercial availability or worldwide representativeness. The shop and tests below are hypothetical.

What the research systems actually produce

TryOnDiffusion, presented at CVPR in 2023, generates a view of one person wearing a garment shown on another. Its two-network approach addresses garment detail and pose change together. The paper's supplementary discussion includes failures caused by segmentation and by removal of person information such as tattoos and accessories. Its image-quality experiments should be read as evidence about generated visuals in those tests. [TryOnDiffusion paper](https://arxiv.org/html/2306.08276v1)

IDM-VTON's July 2024 paper version conditions a diffusion model on both higher-level garment information and lower-level visual features. It aims to preserve garment identity while generating the preview. Its stated limitations include losing human attributes, such as tattoos or moles, in masked regions. The authors' findings support inspecting both the clothing and the person, rather than treating a realistic overall impression as sufficient. [IDM-VTON paper, version 3](https://arxiv.org/html/2403.05139v3)

The 2022 Dress Code dataset extends evaluation across upper-body clothing, lower-body clothing and dresses. Its repository describes catalogue image pairs and auxiliary pose and segmentation information. It also states that the dataset is not released to private companies and requires an access agreement. A publicly readable research repository is therefore not permission to use its underlying images commercially. [Dress Code dataset repository](https://github.com/aimagelab/dress-code)

These sources investigate related but non-identical systems. Their scores and sample images cannot be combined into a league table for a retailer's customers. More fundamentally, generating a plausible garment appearance and predicting physical fit require different evidence. A preview should not inherit a sizing claim merely because both appear on the same product page.

Separate three customer questions

An editorial starting point is to divide the experience into three explicit questions. “How might this style look?” can be answered with a labelled visualisation. “What are the product's dimensions and materials?” should link to verified product information. “Which size should I buy?” needs its own supported process, with uncertainty stated. If no sizing process exists, the preview must not imply that one has run.

Consider a fictional shop testing a patterned jacket. The preview may preserve the general silhouette yet move the pocket, simplify a fastening or invent how the hem falls. A customer might interpret each as a product detail. The shop should decide which details must remain exact, which are illustrative and when an output should be withheld entirely. A small label cannot rescue an image that materially misrepresents the item next to an immediate purchase button.

Keep the source product photograph and specifications accessible beside the preview. Give the customer a clear way to return to them. Explain that the image is generated and may alter details; avoid a success message that says the garment “fits”. If another service provides a size suggestion, distinguish its inputs and result from the image-generation result.

Review fidelity as a set of separate failures

A useful local test could include plain fabric, fine stripes, repeated prints, reflective details, layered clothing and partially hidden sleeves. Select a range of body shapes, skin tones, poses and capture conditions with participants' permission. The point is to discover the service's operating boundary, not to claim a small sample represents every shopper.

For each output, reviewers could assess garment identity, product-detail preservation, person preservation and the clarity of the disclosure separately. Look at the original and generated image side by side at the actual display size. A preview that preserves a face but changes the product fails one purpose; a preview that preserves the product but changes a person's visible features fails another. Do not average those failures into an attractive overall score.

Reserve difficult cases for testing after development choices are made. Repeat a subset to check variation between generations. Record the system version and settings with the review, using a retention policy suitable for participant images. If the application cannot produce an acceptable image, show the original garment and a useful explanation. Quietly substituting a generic model changes the service the person asked for.

Make the image journey understandable

Before upload, explain what image is needed, what processing will occur, who receives it and how long it is retained under the actual implemented policy. Do not promise deletion if the storage and downstream processing cannot deliver it. Avoid requesting faces, background details or other people when the chosen workflow does not need them. Give the customer a route to shop without uploading an image.

Rights review also has several parts: the shopper's permission, the retailer's product-image rights, and the licences and agreements for code, model weights and training or evaluation datasets. Approval of one part does not settle the others. Dress Code's access restriction is a concrete reminder to inspect the actual terms of each proposed asset rather than infer rights from a download link.

The potential benefit is a more personal way to explore appearance before purchase. Whether that reduces uncertainty, increases confidence or changes returns is a separate local measurement question. A retailer could begin by testing whether users understand the preview's limits and can still find the size information. Any later outcome study should distinguish visual appeal from correct product expectations.

The first release decision should therefore be narrow: which garments, photographs and customer tasks have acceptable visual fidelity, clear rights and understandable limits? A generated image can be useful without being a measurement. Preserving that distinction is what makes the preview an honest shopping tool.