We build AI visualization into commerce and consumer products: virtual try-on for clothing, footwear, eyewear and accessories; furniture and decor placed into a photo of a real room; and on-model catalogue imagery generated without a photoshoot. Delivered as an API, an embedded feature, or a full build handed over with the source code.
Exelero builds computer vision and generative imagery into production products. Work in this category includes virtual try-on, where a customer photo and a catalogue product image produce a realistic image of that product on that person; product placement, where furniture, decor, appliances or fixtures are rendered into a photograph of a customer’s own space; generated on-model imagery, producing catalogue photography across body types without a studio; and conventional computer vision such as detection, segmentation, OCR and document extraction.
Builds include the production layer that pilots usually skip — confidence scoring and refusal, evaluation sets, cost-per-generation modelling, latency budgets and privacy handling for customer photographs. Delivery can be an API, an embedded feature inside an existing product, or a full build handed over to the client.
Online returns ran at 19.3% of ecommerce sales in the National Retail Federation’s 2025 returns study. Apparel runs higher — 20–40% depending on category — and fit and sizing is the single largest cause, responsible for roughly half of apparel returns. Every one of those is a shipment out, a shipment back, a restock, a refund and often an item that never sells again at full price.
The same gap exists in furniture, where the question is whether the sofa fits the room and matches what’s already in it, and in any category where the buying decision depends on how something looks in a context the retailer can’t photograph.
Clothing, footwear, eyewear, jewellery, watches and accessories rendered onto a customer’s photo, working from standard catalogue images with no 3D modelling and no body measurements.
Furniture, decor, appliances, fixtures, flooring and paint rendered into a photograph of the customer’s own room at estimated scale and matched lighting.
One garment shown across a range of body types and skin tones without a studio booking, turning a photography budget into a per-image cost and making size-inclusive merchandising practical.
Detection and tracking, segmentation, classification, OCR and document extraction, quality inspection, and edge deployment on constrained devices.
Confidence scoring with automatic refusal below a configurable threshold; fallback to original imagery; product fidelity controls so colour, pattern and cut survive generation; input quality scoring with guidance back to the user; evaluation harnesses and regression testing; cost-per-generation modelling; caching; latency budgeting; watermarking and generated-content metadata where regulation requires it.
Processing without biometric template storage, configurable retention down to immediate deletion, consent capture, automated moderation on uploads, regional processing and self-hosted deployment where images cannot leave the client’s environment.
This is the constraint that shapes every serious build in this category. A shopper who sees a warped garment, a wrong colour or a print in the wrong place doesn’t conclude the feature is imperfect — they conclude the retailer is careless. The failure mode is asymmetric: a good result adds modest confidence, a bad one costs a customer.
So the system can’t be optimized for average output quality. It has to be optimized for the floor. We build the confidence scoring and the refusal path before tuning quality, on the assumption that bad outputs never reach zero — which inverts the usual roadmap and makes the feature shippable much earlier. The second decision that matters is cost. Inference cost moves by roughly an order of magnitude between a naive implementation and a considered one, which is the difference between a feature viable across a full catalogue and one that only works as a demo.
We build in 8–12 weeks from scope to launch, from foundations already built, with scope fixed in the first two weeks. An API integration is faster — often a matter of weeks. The full window applies when visualization is being built into a product, with evaluation, tuning against your catalogue and the confidence-threshold work included.
Two to three weeks. Can this work on your catalogue, how well, at what cost per generation, and what’s the failure mode? Ends with real output on your own products and an honest recommendation, including when that recommendation is not to build it.
Delivered as an API, embedded into your product, or as a full white-label build.
Source code, infrastructure and documentation where the engagement is a build.
Model updates, evaluation maintenance and cost optimization on a retainer.
Fashion, footwear and accessory retailers and marketplaces. Furniture, decor and home improvement retailers. Eyewear, jewellery and cosmetics brands. Resale and rental platforms. Ecommerce platforms and agencies embedding visualization for their own clients. Catalogue and content teams producing imagery at scale. Any product where the buying decision depends on how something looks in a context you can’t photograph.
8–12 weeks from scope to launch for a full build; an API integration is faster. A feasibility sprint comes first, so you see real output on your own products before committing.
A customer photo and a product image go in. The system segments the person or the room, estimates pose, scale and lighting, and generates a new image with the product placed realistically, preserving its actual colour, pattern and cut. No 3D model of the product and no body measurements are required.
The data says it targets the right problem. Fit and sizing causes roughly half of apparel returns, so try-on addresses the largest driver directly rather than a peripheral one. On measured outcomes, ASOS reported a 160-basis-point reduction in its returns rate following a programme that included virtual try-on. Because results vary by category, price point and market, we build A/B testing into the integration so you get your own number on your own catalogue.
An ordinary phone photo. Full-body works best for clothing, half-body and selfies for tops, eyewear and cosmetics, a straight-on shot for rooms. Input quality is scored automatically with specific guidance when a retake would help.
Yes — furniture, decor, appliances, flooring and fixtures placed into a photo of the customer’s own room.
Whatever you configure, including immediate deletion after generation. No biometric templates are created or stored, consent capture is built in, and regional or self-hosted processing is available where data residency is required.
Yes. Self-hosted and on-premise deployment is standard where data can’t leave your environment.
The fastest way to evaluate this is on your own catalogue. Give us a handful of product images and we’ll generate results you can judge for yourself.