Fashion has two image-heavy bottlenecks: designers need many visual variants early, and shoppers cannot see how a garment looks on them, which drives returns. Generative image models now produce design concepts from sketches and prompts, and "virtual try-on" (VTON) models drape a product photo onto a photo of a person or a model. Used for design ideation, catalogue imagery and shopper-facing try-on, with human review and clear labelling of synthetic images.
Problem & value
- Today: collections are visualised through sketches, samples and photo shoots with models, studios and retouching. Online shoppers rely on flat product photos and size charts; fashion e-commerce has high return rates, and sizing and "does it suit me" are commonly cited reasons.
- AI value: (1) faster concept exploration and colourway/print variants before sampling; (2) on-model imagery for the whole assortment without a shoot per SKU, across body types; (3) shopper try-on to raise confidence and reduce fit-related returns.
- Evidence: vendors publish conversion and return-rate improvements, but these are self-reported and vary by category; run your own A/B test before budgeting on them. Realism of fabric drape, fit and sizing is the weak point: try-on shows appearance, not physically accurate fit (see REC-012 for size recommendation).
Variants
- Designer ideation: sketch-to-image, print/pattern and colourway generation, moodboards; 3D garment tools (CLO, Browzwear) for technically valid patterns.
- Catalogue / on-model imagery: swap garments onto studio models or synthetic models (see GEN-001).
- Shopper try-on (B2C): customer uploads a photo or uses an avatar on a shop or app.
- Marketplace / seller tools: sellers turn flat-lay photos into on-model images.
- Private use: consumer apps to preview outfits from one's own wardrobe.
- Out of scope here: identity-preserving portraits (GEN-026), body measurement for sizing (REC-012).
How it works
- garment image (flat/ghost/on-model) + person image (or generated model)
- preprocessing: person parsing/segmentation, pose estimation, garment mask
- try-on generation (diffusion/transformer: garment conditioning + inpainting of the clothing region)
- post-processing: upscaling, artefact check, background/colour consistency
- quality gate (automatic checks + human review for catalogue use)
- labelling (synthetic-content marking/metadata)
- shop / PIM / DAM / app
- Core models: diffusion- or flow-based image models conditioned on garment features; modern variants use image-editing foundation models that take two reference images (person + garment).
- Design side: text/sketch-to-image with style or LoRA adaptation to the brand's look.
- Human-in-the-loop: designers select and refine concepts; merchandisers approve catalogue images (wrong print, missing details, altered logos); shopper try-on is explicitly shown as a preview.
Data preparation
Needed data are clean garment product photos (flat-lay, ghost mannequin or on-model), SKU master data with size, colour and category, and, for try-on, customer or model photos. Product images typically live in the shop system, a PIM or DAM, ERP article master data and on photographers' drives; customer uploads arrive through the web shop. Typical gaps are inconsistent backgrounds and angles, missing or wrong attributes, duplicates across colourways, unclear model releases and image rights, and customer photos that count as personal data.
Preparation steps: standardise photo shooting guidelines (background, light, resolution), clean article attributes in the PIM (PREP-011, PREP-010), assemble a gold-standard set of approved on-model images for quality checks (PREP-016, PREP-018), document model releases and rights, and define retention rules so customer photos are deleted right after the session (PREP-025). Background removal and tagging can be semi-automated, see GEN-011.
Effort share is medium, about 25 to 35 percent, mainly for image standardisation, attribute cleanup and the legal documentation of image and customer-photo handling.
Implementation paths
Which path for whom
- SMB: A, because mature EU-capable try-on and catalogue-image products exist as shop plug-ins and the process is standard.
- Enterprise: B, because try-on is a customer-facing core feature at high volume, so API costs and brand control justify integration into the own shop.
- Developer: C, because open try-on models can be run and fine-tuned on own GPUs for experiments and custom pipelines.
- Private: A, because shopping apps and consumer try-on tools need no setup.
Path A — Buy (off-the-shelf)
- Shopper try-on / catalogue plugins: specialised vendors and services, e.g. Botika, Vue.ai, Style3D, Kivisense, Pixelcut/Photoroom (product imagery), plus Google's consumer try-on features in Search/Shopping. EU-based options include Botika (Spain); verify the processing region for each.
- Design tools: Adobe Firefly (in Photoshop/Illustrator), Midjourney, Vizcom, Browzwear and CLO (3D garment design and digital samples), Canva AI.
- Choose A when: assortment is moderate, no custom shopper-facing experience needed. Check EU hosting, deletion of uploaded customer photos, no training on your data, and Art. 50 labelling support.
Path B — Build on cloud APIs
- Dedicated try-on API: Google Vertex AI "Virtual Try-On" model (person image + garment image, EU regions available for some Vertex features, verify) [1].
- Image-editing models: multi-image editing in Gemini image models, GPT-image models, and FLUX Kontext (Black Forest Labs, Germany) via API for garment swaps and design variation [2].
- Integration: Shopify/Shopware/Magento plugin, PIM/DAM connector, queue + review UI, CDN.
- Choose B when: you need try-on inside your own shop/app or batch production for a large catalogue.
Path C — Self-host (open source)
- Try-on models: CatVTON [3], IDM-VTON [4], and related research models; general image models: FLUX.1 (dev/Kontext dev; non-commercial licence for dev weights, check terms), Stable Diffusion family, Qwen-Image-Edit; serving via ComfyUI or diffusers.
- Segmentation/pose: SAM-family models, OpenPose/DWPose-class estimators, human parsing models.
- Hardware: HW2 runs lightweight try-on models at several seconds per image; HW3 for larger editing models, high resolution and batch. Check each weight licence: many research try-on weights are non-commercial and trained on datasets with restrictive terms.
- Choose C when: customer photos must stay on-prem/in the EU, or volumes make per-image fees expensive.
Path D — Custom training
Worth it for a distinctive brand look (LoRA/fine-tune on your own photography), unusual garment categories (bridal, technical wear, jewellery, shoes) or consistent house models. Needs a few hundred to thousands of paired garment/on-model photos with clear usage rights and model-release consent.
State of the art (2026)
- Try-on quality has moved from research demos to productised APIs and image-editing models that accept a person and garment image with decent texture, logo and print retention on simple garments.
- Open research (CatVTON, IDM-VTON and successors) gives usable results on upper-body garments and dresses; benchmarks such as VITON-HD and DressCode are used, but they are limited in pose and body diversity.
- Failure modes: distorted prints and text, wrong sleeve/hem length, loss of fine details (lace, pleats), unrealistic drape of heavy or sheer fabrics, layered outfits, accessories and shoes, extreme poses and plus-size or non-standard bodies (bias and quality gaps), changed skin tone or body shape.
- Design side: generative tools are strong for ideation but weak for producible patterns; 3D pattern software with physics-based cloth simulation is still needed for fit and production.
- Video try-on and 3D avatar try-on are emerging, with less reliability.
Shared resources & combinations
- Same image stack as GEN-001 (product imagery), GEN-011 (retouching, upscaling) and GEN-018 (staging); one GPU server (HW3) or one API contract can serve all.
- Combines with: REC-012 (size & fit, to convert try-on into the right size), PRED-034 (visual trend forecasting feeds design briefs), GEN-008 (3D assets), GEN-004 (layout for lookbooks), GEN-026 (portraits), ANOM-033 (returns analytics to measure impact).
Effort, cost & KPIs
- Path A: days to weeks; typically per-image or per-SKU subscription pricing.
- Path B: 1–3 months for a shop-integrated flow incl. review UI; cost is per generated image plus engineering.
- Path C: 2–4 months; GPU server, ComfyUI/diffusers pipeline, licence review, ops.
- Path D: adds data collection, rights clearance and evaluation; several months.
- KPIs: try-on usage rate, conversion uplift (A/B), return rate for fit/appearance reasons, cost per catalogue image vs. shoot, time from sketch to approved concept, share of generated images rejected in review, complaints about misrepresentation.
People & change
Involve management and the e-commerce or design lead from the start; they own the business case and the brand rules. IT integrates the shop and handles accounts. The DPO is essential because customer photos are personal data and body images can be sensitive: clarify legal basis (consent), AVV with the vendor, deletion and, where scale or new technology applies, a DPIA (Art. 35 GDPR, GUIDE-005). Legal checks image rights, model releases and AI labelling (Art. 50 AI Act).
Works council: depends. Customer-facing try-on does not monitor staff, but internal design and content tools with per-user logging can fall under §87(1) no. 6 BetrVG, and the work-process change must be notified under §90(1) no. 3; ask the council early if one exists. AI literacy: support photographers, designers and shop staff with a short training on limits, bias in body shapes and labelling (Art. 4 AI Act, GUIDE-006). Typical concerns are job loss for photographers and models and unrealistic fit: start with a pilot on one product category, compare against real photos, state clearly that the try-on is an approximation, and keep a human approval step for catalogue imagery (GUIDE-008, GUIDE-009).
Risks, regulation & compliance
- EU AI Act: generally limited risk. Art. 50 requires machine-readable marking of AI-generated images by providers, and disclosure for deepfakes by deployers; realistic synthetic models and edited photos of real people should be labelled. Not prohibited by Art. 5 and not high-risk, but do not use try-on for emotion recognition or biometric categorisation of shoppers (inferring sensitive traits from body images would fall under Art. 5/Annex III concerns). AI literacy (Art. 4) applies [5][6].
- GDPR: customer photos and body shape data are personal data; images can reveal health or ethnic origin (special categories, Art. 9 risk). Use consent or clear contractual basis, purpose limitation, short retention (delete after the session), no use for model training without consent, DPIA for large-scale B2C try-on, processor agreements and EU/EEA processing.
- Likeness and model rights: models' contracts must cover digital replicas; synthetic models must not resemble real persons; copyright of designs and training data (AI Act Art. 53 duties on general-purpose model providers); trademark/logo handling.
- Consumer law: images must not mislead about colour, fit or material (unfair commercial practices rules); show a notice that the preview is illustrative.
- Bias and body image: test across skin tones, sizes and ages; avoid automatic body slimming.
- Mitigations: human approval of catalogue images, C2PA/IPTC provenance metadata, visible "AI-generated" tags, upload-and-delete flows, red-teaming for misuse (e.g. try-on on other people's photos, non-consensual imagery) with upload checks and content filters.
Getting started
- Pick one category (e.g. tops/dresses) and baseline return rate and photo cost per SKU.
- Test 2–3 Path A tools and one API with 50 of your own garments, including print-heavy ones; score fidelity.
- Define the review workflow and labelling (metadata plus visible tag) with legal/DPO input.
- Pilot shopper try-on for one category with an A/B test and a delete-after-session policy.
- Decide on B/C only if volume, data-residency or brand-consistency needs justify engineering.
Sources
- Google Cloud — Virtual Try-On on Vertex AI — https://cloud.google.com/vertex-ai/generative-ai/docs/image/generate-virtual-try-on-images
- Black Forest Labs — FLUX repository and FLUX.1 Kontext — https://github.com/black-forest-labs/flux
- CatVTON — https://github.com/Zheng-Chong/CatVTON
- IDM-VTON — https://github.com/yisol/IDM-VTON
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- AI Act Article 50 — transparency obligations — https://artificialintelligenceact.eu/article/50/
- CLO Virtual Fashion — https://www.clo3d.com/
- Browzwear — https://browzwear.com/