Product ML Engineer

Vestiaire Collective boosts sales 5% with Photoroom standardised product shots. 13 Oct 2026 Photoroom have published results from its work with Vestiaire Collective, the global platform for pre-loved luxury fashion. Vestiaire Collective split its new listings into two random groups. One had its main photo edited by Photoroom, the other by the photo clipping tool Vestiaire Collective built in-house years ago. Listings in the Photoroom group sold at a 5% higher rate of liquidity. Their product pages drew 3 to 7% more unique views, with bags and shoes gaining the most. Vestiaire Collective chose Photoroom after testing several generative AI tools and finding that every one of them changed the items. A photo that changes the item breaks two things on a resale platform. Vestiaire Collective authenticates a share of items from their photos alone, checking labels, stitching and how light falls on the material. Buyers can also return anything that turns out to be fake or has a fault the seller didn't disclose. Smooth out a scuff or redraw a logo and both are at risk. Photoroom's own testing points the same way. In a benchmark of 4,250 AI-generated images across 850 real products, the best AI image-editing models on the market passed a full product fidelity check only 29% of the time. Buyers notice too: 63% of UK consumers say inconsistent product images or branding make a marketplace less trustworthy, according to Photoroom's State of GenAI in Marketplaces 2026 report. We tested different GenAI solutions, and each time they changed some details. Then it was no longer possible to authenticate from the pictures or to make sure buyers could trust them. Photoroom maintained product fidelity, and that's clearly the reason we chose it. - Thomas Hezard, Chief Product Officer, Vestiaire Collective Vestiaire Collective built its own photo clipping tool years ago, because nothing on the market was accurate enough. Photoroom now runs that work as a service for enterprises. Its Visual Agents system, launched in August 2026, brings together the right AI models and quality checks for each catalogue, so every image that goes live is true to the product and ready to sell. Vestiaire Collective operates in 70 countries and adds around 40,000 listings a day. Sellers range from private sellers photographing items at home to semi-professional sellers using AI tools to enhance their photos. Sellers list the way they always have. Before a listing goes live, the Photoroom API removes the background from the main photo, adds a shadow, centres the item and corrects the lighting. Bags and shoes get a deeper shadow to show their shape, clothing laid flat gets little or none. It took eight weeks to go from deciding to test to having results. The same photos run in Vestiaire Collective's performance marketing ads, which became 2% more efficient, while digital authentication was unaffected. Photoroom's pipeline for Vestiaire Collective combines its own models with open models, including FLUX.2 [klein] from Black Forest Labs and DINOv2 from Meta. It runs on NVIDIA H100 and H200 GPUs, with GPU capacity from Nebius and inference optimised using NVIDIA TensorRT and NVIDIA Triton Inference Server. Some of Photoroom's models only get a few requests an hour, so they shut down when idle and restart when traffic arrives. As a member of the NVIDIA Inception programme, Photoroom worked with NVIDIA to put NVIDIA Dynamo Snapshot into production. A new GPU server now comes online from around three and a half minutes to about a minute, so a spike in listings doesn't slow anything down. We looked at how first-hand players create emotion to convince a buyer to purchase a €1,000-plus item. Then we looked at our photos, and the gap was clear. - Thomas Hezard, Chief Product Officer, Vestiaire Collective Nobody returns a photo. They return the product. Vestiaire Collective tried the tools that change products and picked the one that doesn't, and more of its items now sell. Marketplaces have spent years building their own tools to get this right, and we want them to be able to buy it instead. - Matt Rouif, Co-founder and CEO, Photoroom.

Ten years after shoppable posts changed how Retail Times buy, AI is set to change how Retail Times sell. Reading Time: 4 mins read By Matt Rouif, CEO and co-founder of Photoroom When shoppable posts first appeared around a decade ago, they helped close the gap between discovering a product on social media and being able to buy it. That model now feels normal across TikTok Shop, creator storefronts and platforms such as Vinted, but as AI creates more routes to discovery, the pressure is shifting further upstream. Retailers increasingly need the right imagery and product information ready for every environment in which customers might find them. The bottleneck has moved While shoppable posts made it easier to connect an image directly to a transaction, the decade since has exposed a much broader problem around scale. Retailers now have more routes to market than ever before, but maintaining the content needed to support those channels becomes a significant production challenge when the same product has to appear across marketplaces, social platforms and product pages, often with different requirements in each place. For smaller retailers, that can still mean stitching together several tools simply to get one product ready to sell. A business might use remove.bg for backgrounds, Canva for design work and Photoshop for more detailed editing, before moving into a separate AI tool for descriptions or metadata. Each platform can solve one part of the process, but the friction becomes obvious when products have to be moved between them repeatedly and the same workflow needs to be replicated across hundreds of listings. What Retail Times hear repeatedly from smaller sellers is that the problem is rarely a lack of ambition. It is the amount of manual work that sits between having a product and getting that product live in a form that is ready to sell, particularly when every additional sales channel introduces another set of requirements around how the product needs to look and how quickly that content needs to be produced. AI can change the economics of visual production For much of ecommerce's development, increasing visual output meant increasing production resource alongside it. More products and more markets brought additional photography and editing work, which placed a natural limit on how quickly smaller businesses could scale and made it difficult to expand without adding significant creative capacity. Generative AI changes that equation by allowing content to be created or adapted much closer to the point at which it is needed. For smaller businesses, the opportunity is therefore less about making one image faster and more about expanding what they are able to produce without requiring the size and cost of the creative operation to grow at the same rate as the commercial footprint. That was one of the reasons Retail Times focused on embedding Photoroom directly into Shopify rather than building another tool that sits alongside the store. Through the integration, a retailer can import its catalogue and batch-edit up to 250 products, while also generating missing product angles from an existing image and creating SEO metadata in the store's own language before publishing the changes back into Shopify when they are ready. The point is to remove unnecessary movement between disconnected workflows, because the value of the technology is reduced if a seller still has to take the same product through several separate systems before it can go live. That only works if accuracy remains the foundation of the process. In commerce, the objective is not simply to create an attractive image, because the image also has to give the customer an accurate expectation of the product they will receive. Generative models have become extremely capable at producing visually convincing content, but ecommerce imposes a different standard, where relatively small changes to colour or texture can make an image less faithful to the product even if the result still looks polished. The product itself therefore has to remain the fixed point, with creative flexibility built around it rather than replacing it. AI can give retailers much more freedom in the way products are presented, but that freedom only becomes useful when the factual attributes of the product remain consistent throughout the process. If imagery creates an expectation that the physical product does not meet, the consequences can quickly appear through returns and lost trust. Speed and fidelity therefore have to be measured together, particularly as AI-powered search and shopping assistants begin surfacing products in environments that retailers do not fully control. Looking at generative AI purely as a cheaper alternative to photography misses much of the larger opportunity. Lower production costs can clearly make experimentation easier, but the more significant shift is towards treating product visuals as a scalable production layer that supports the wider commerce operation rather than as something that has to be rebuilt each time a new campaign or channel appears. Instead of recreating content from scratch whenever a product needs to appear in another marketplace or format, retailers can adapt visual assets around the same underlying product information and brand standards. That turns imagery into something much closer to an operational capability, because the same visual system can continue supporting products as they move across different channels and markets. Retail Times is already seeing what that can mean at very different ends of the market; Decathlon for example, has used Photoroom to reduce image production costs by 99% at catalogue scale, while in the UK, fashion retailer Layer increased online sales by 50% and Ameliora increased Shopify conversions by 30%. In each case, the value comes from creating a repeatable way of producing the content required to sell, rather than simply generating a larger volume of imagery. That matters because ecommerce is becoming less centred around a single destination. A retailer may now need the same product to appear convincingly on its own store and across a marketplace, while also adapting it for social commerce or an AI-powered shopping experience. The businesses that benefit most will not be those producing the greatest volume of content, but those that build visual systems capable of keeping pace with modern commerce while preserving the accuracy and consistency that customers still depend on.