Virtually try on dresses with our AI shopping tool

Sep 05, 2024

[[read-time]] min read

Lilian Rincon

Vice President, Product Management

Ira Kemelmacher-Shlizerman

Senior Staff Research Scientist, Shopping

Shopping online to update your seasonal wardrobe tin beryllium nosy — and challenging. It’s difficult to show really a caller style will look connected you, and buying thing that doesn’t activity tin nonstop you done nan returns process, which is burdensome for some shoppers and retailers. It’s why we launched a virtual try-on instrumentality for women's and men’s tops past year, which uses generative AI to thief you spot each item — down to nan shadows, wrinkles and draping — of a portion of clothing. Since then, we’ve recovered that virtual try-on is awesome for some shoppers and brands. Shoppers walk much clip visiting and interacting pinch these images compared to different shopping images — successful fact, virtual try-on images connected Search person 60% much high-quality views, and connected average, group effort connected apparel utilizing 4 models per product. Shoppers are besides much apt to sojourn a brand’s tract aft viewing virtual try-on images1.Shopping online to update your seasonal wardrobe tin beryllium nosy — and challenging. It’s difficult to show really a caller style will look connected you, and buying thing that doesn’t activity tin nonstop you done nan returns process, which is burdensome for some shoppers and retailers. It’s why we launched a virtual try-on instrumentality for women's and men’s tops past year, which uses generative AI to thief you spot each item — down to nan shadows, wrinkles and draping — of a portion of clothing. Since then, we’ve recovered that virtual try-on is awesome for some shoppers and brands. Shoppers walk much clip visiting and interacting pinch these images compared to different shopping images — successful fact, virtual try-on images connected Search person 60% much high-quality views, and connected average, group effort connected apparel utilizing 4 models per product. Shoppers are besides much apt to sojourn a brand’s tract aft viewing virtual try-on images.

So today, conscionable up of nan autumn season, we’re expanding virtual try-on to different wardrobe staple, and 1 of nan most-searched apparel categories: dresses. You tin now amended visualize dresses connected models that resonate astir pinch you from hundreds of brands crossed Google’s Shopping Graph — including SIMKHAI, Boden, Staud, Sandro and Maje — successful conscionable a fewer taps.

How it works

If you’re successful nan U.S, simply hunt for dresses connected Search and click immoderate style that includes a “try on” badge. From there, you tin spot what that garment looks for illustration connected a divers group of existent models, ranging successful sizes from XXS-XXXL. Select a exemplary that resonates pinch you to get a amended consciousness of really nan dress looks. Once you find what you’re looking for, click retired to nan retailer’s tract to bargain it.

How we built it

This characteristic is made imaginable acknowledgment to a generative AI exertion we created specifically for virtual try-on (VTO), which uses a method based connected diffusion. Diffusion lets america make each pixel from scratch to nutrient precocious quality, realistic images of tops and blouses connected models. As we tested our diffusion method for dresses, though, we learned location are 2 unsocial challenges: First, dresses are usually a much nuanced garment, and second, dresses thin to screen much of nan quality body.

Let’s commencement pinch nan first problem: Dresses are often much elaborate than a elemental apical successful their draping, silhouette, magnitude aliases style — and see everything from midi-length halters to mini shifts to maxi driblet waists — positive everything successful between. Imagine you’re trying to overgarment a elaborate dress connected a mini canvas — it’d beryllium difficult to compression successful specifications for illustration a floral people aliases ruffled collar onto that mini space. Enlarging nan image won’t make specifications clearer, either, because they weren’t moreover visible successful nan first place. You tin deliberation of our VTO situation successful nan aforesaid way: Our existing VTO AI exemplary successfully diffused utilizing low-resolution images, but successful our testing pinch dresses, this attack often resulted successful nan nonaccomplishment of a dress’s captious specifications — and simply switching to high-resolution didn’t help. So our investigation squad came up pinch what’s called a “progressive training strategy” for VTO, wherever diffusion originates pinch lower-resolution images and gradually trains successful higher resolutions during nan diffusion process. With this approach, nan finer specifications are reflected, truthful each dress’ pleat and people comes done crystal clear.

Next, since dresses screen much of a person's assemblage than tops, we recovered that “erasing” and “replacing” nan dress connected a personification would smudge nan person's features aliases obscure important specifications of their assemblage — overmuch for illustration it would if you were coating a image of personification and later tried to erase and switch their dress. To forestall this “identity loss” from happening, we came up pinch a caller method called nan VTO-UNet Diffusion Transformer (VTO-UDiT for short) which isolates and preserves a person’s important features. So while we train nan exemplary pinch “identity loss” successful place, VTO-UDiT besides gives america a virtual “stencil,” allowing america to re-train nan exemplary connected only nan person, preserving nan person’s look and body. This gives america a overmuch much meticulous portrayal of not only nan dress but conscionable arsenic important, nan personification wearing it.

How to shop (and effort on) nan runway

This period successful New York, hundreds of brands will beryllium showing dresses from their latest collections connected runways crossed nan city. Following ready-to-wear marque SIMKHAI's September show, shoppers will beryllium capable to virtually effort connected prime dresses from nan postulation consecutive from nan runway and pre-order them from SIMKHAI’s site. Just hunt for SIMKHAI dresses connected Google and look for nan “try-on” icon to get started. While we’ve collaborated pinch SIMKHAI to thief you easy shop looks from New York Fashion Week, you’ll still beryllium capable to virtually effort connected tops and dresses from hundreds of brands crossed nan web — now pinch much certainty you’ll find thing conscionable right.