Virtual try-on for ecommerce: a 50-SKU benchmark of garment fidelity, sizing cues, logo preservation, and on-brand campaign output
A 50-garment prompt test measured generation time and cost, not fit or fidelity scores. Use a merchant-controlled scorecard before treating virtual try-on as fit guidance.
Lamina Team
Product Team @ Lamina
Can this 50-SKU virtual try-on benchmark identify the best ecommerce platform?
No. The 50-garment experiment tracks image-generation cost and latency across prompt variants, yet reports no scored garment fidelity, sizing cues, logo preservation, campaign usability, acceptance rate, or retry rate. Calling any platform a universal winner from this record would be fiction.
The conclusion is narrower: treat virtual try-on as two linked, separate ecommerce jobs. Generative try-on makes on-model PDP and campaign images. Fit prediction tells a shopper whether a specific size is likely to work on their body. A render can nail the first and prove nothing about the second.
That split protects conversion and trust. Use generated images to show styling, drape, texture, silhouette, and brand direction. Then back them with measurement-based size guidance, product-specific fit notes, and plain uncertainty where fit evidence is thin.
Lamina prompt experiment on a fixed set of 50 ecommerce garments. The supplied record contains cost and latency observations only; it does not contain quality scoring, reviewer acceptance data, or fit-validation outcomes.
Generation cost per image
over Experiment recorded 2026-08-04
Average generation latency
over Experiment recorded 2026-08-04
Garment fidelity, sizing cues, logo preservation, and campaign usability
over Experiment recorded 2026-08-04
What did the prompt experiment actually prove?
It showed an iteration-time trade-off, not a quality winner. More detailed garment and reference instructions took longer to generate, while the recorded per-image generation charge stayed constant.
That matters to a creative team. A fidelity-first brief or reference-heavy campaign brief may need a bigger review window; a generic brief gets you into rough exploration faster. None of those figures includes human art direction, approval rounds, retouching, revisions, or media spend.
The missing test condition matters too. The record never gives the number of runs, leaving these as observations from one experiment rather than a production guarantee. It offers no evidence that the extra time improved logos, fabric detail, sizing signals, or campaign output.
| Metric | Value | Source |
|---|---|---|
| Models in one feasibility evaluation — evidence of model trade-offs rather than a single universal visual-fidelity winner | 5 | doi.orgas of 2025-11-28 |
| Participants in a VTON user study — shoppers wanted fit accuracy and reliability cues, so visual output alone is not enough | 24 | dl.acm.orgas of 2026-01-01 |
| Fit-accuracy improvement reported by the TrustFit paper — promising, but a single-study result that needs merchant validation | 12–15% | doi.orgas of 2026-05-06 |
| Fashion categories claimed by Tstars-Tryon 1.0 — broad category coverage is an author claim, not independent proof of ecommerce readiness | 8 | ar5iv.labs.arxiv.orgas of 2026-04-01 |
Can virtual try-on tell shoppers whether clothing will fit?
Virtual try-on can show how a garment looks on a body. By itself, it cannot guarantee physical fit or size. Google describes its image-based apparel Try-On as a useful vibe check, while current fit-focused research still treats body–garment alignment and fitting fidelity as an open technical problem.
Do not turn a polished image into a sizing promise. Shoppers need evidence tied to the product’s actual size pattern, textile behavior, body measurements, and, preferably, observed return outcomes. Put a visual render alongside those signals; do not let it stand in for them.
There is a real commercial reason to surface fit information. Randomized field experiments found that fit information raised conversion and order value while cutting fulfillment costs tied to returns and multi-size home try-on. Fit validation deserves its own workstream.
Which virtual try-on models are credible candidates for garment-detail testing?
Leffa and FitDiT are credible places to start visual garment-fidelity testing. Neither has earned the title of best production platform across every apparel catalog. In a five-model feasibility evaluation using VITON-HD and DressCode, they delivered the strongest visual-fidelity results; CatVTON stressed throughput and affordability, while IDM-VTON offered optimization flexibility.
For printed tees, striped knits, type-heavy garments, and detail-dense lines, put FitDiT through a dedicated stress test. Its design targets texture-aware preservation and size-aware fitting, including high-detail attributes such as stripes, patterns, and text.
Logo-critical products deserve harsher scrutiny than a generic fashion sample. IDM-VTON was designed to preserve garment identity and low-level details. Newer preprints, including Oxygen-TryOn, claim preservation of logos, prints, textures, and silhouettes; that earns them a test, not permission to publish output without review.
How should a merchant run a 50-SKU virtual try-on benchmark?
Build a SKU set around the ways your catalog breaks
Use the same garments for every candidate: solid basics, stripes, dense prints, visible wordmarks, embroidery, reflective trim, layered looks, and silhouette-sensitive pieces. Include the categories and product-detail constraints that shape your PDPs. Skip the handpicked easy images.
Keep visual production separate from size validation
Score garment shape, hem and sleeve geometry, texture, print alignment, logo legibility, embroidery, body and pose integrity, and campaign-style adherence as image-quality criteria. Test sizing claims separately, using actual garment measurements, size patterns, textile characteristics, shopper feedback, and returns data.
Use a blinded review scorecard before campaign release
Have brand, merchandising, and ecommerce reviewers judge outputs without knowing which tool or prompt made them. Log pass, revise, or fail, plus the reason for each call. Logo placement, typography drift, altered color, and misleading fit cues need explicit failure codes.
Measure production economics after review, not generation
Track generation time and model charges, then add retries, human selection, art direction, approval, and downstream edits. A low-cost image stops being low-cost once it keeps missing a wordmark, fabric construction, or the brand’s styling brief.
Publish only the claim your evidence can carry
Use approved virtual try-on output for on-model merchandising and campaign creative. Where your fit study has not validated a size outcome, call the experience visual styling guidance. Keep the size chart, fit notes, and measurement-based recommendation in view.
How do you turn virtual try-on into on-brand campaign creative without weakening product truth?
Treat the garment reference and campaign brief as separate controls. The garment reference locks product geometry, material behavior, color, print placement, embroidery, and logo position. The art-direction brief sets model casting, pose, lighting, background, crop, composition, and the campaign’s permitted visual language.
A reference-heavy brief is not automatically stronger. It may protect silhouette and brand marks, yet the supplied experiment never measured whether it improved creative quality. Review it against a defined style scorecard instead of assuming more references make better work.
Human review is still the release gate. Use AI generation for new concepts, complex styling, reusable on-model creative, and material detail at production pace. Put a brand-aware art director in charge of approving the frames customers actually see.
What should ecommerce teams buy after a virtual try-on pilot?
Buy the platform that passes your SKU scorecard and fits the job at hand. For production on-model imagery, assess garment identity, controllable art direction, repeatability, throughput, and review burden. For customer size guidance, require independently validated measurement and outcome evidence.
Do not buy a fit claim because a model image looks physically plausible. FitVTON notes that many diffusion-based try-on systems prioritize texture preservation and can create plausible images that fail to reflect authentic fit across body shapes. Clear confidence messaging is a product requirement.
The strongest deployment pairs virtual try-on for believable, on-brand visual merchandising with a measurement-led fit layer for size decisions. That setup scales fashion creative without making a render carry evidence it does not have.
Methodology
Original Lamina experiment run 2026-08-04. Hypothesis: For a fixed set of 50 ecommerce garments, a structured virtual-try-on prompt that explicitly anchors garment geometry, material behavior, size cues, logo placement, and brand art direction will outperform a generic try-on prompt on product fidelity and campaign usability, while a reference-heavy prompt will further improve logo and silhouette preservation but may reduce creative campaign quality.. Measured 3 variant(s) for cost and latency on the Lamina image engine; numbers cited here are our own measurements.
Continue reading
Data report: Ecommerce virtual try-on benchmark — measuring garment fidelity, model consistency, and brand accuracy across 100 catalog images
A 100-image virtual try-on benchmark should separate garment fidelity, model consistency, and brand accuracy—and report failures, not just attractive averages.
Lamina Team
Product Team @ Lamina
Virtual try-on for ecommerce: turning digital dressing rooms into on-brand product and campaign creative
Virtual try-on can help shoppers evaluate products on a PDP and give brands a reusable, on-brand creative mechanic for campaigns, email and events.
Lamina Team
Product Team @ Lamina