Google announced Thursday that it expanded its generative AI-powered virtual try-on tool to support dresses, allowing users to virtually wear thousands of dresses from hundreds of brands, including Boden, Maje, Sandro, Simkhai and Staud.
According to the company, dresses were one of the most searched apparel categories for the tool. However, as Google explained in today’s blog post, its current diffusion technique is challenging to use with dresses, as they are more detailed and complex compared to other clothing items.
To provide more context, Google Shopping released the tool last year, using its own diffusion technology to create high-quality, lifelike images of tops and blouses. It simulates how the clothing would drape, fold, cling, and form wrinkles and shadows on real people in various poses.
Due to the intricate details of dresses, the existing diffusion model struggled to accurately capture detailed dress prints such as floral or geometric patterns. While the model could handle low-resolution images, a different method was needed for dresses to avoid losing important details. To address this issue, Google said it developed a new training strategy that starts with lower-resolution images and gradually incorporates higher resolutions.
Additionally, as dresses typically cover most of the body and come in various lengths (such as midi, maxi, and mini), placing a virtual dress on a person often leads to the obscuring or blurring of body details. A new technique called the VTO-UNet Diffusion Transformer (VTO-UDiT) aims to solve this problem by preserving a person’s features while erasing and replacing the dress, resulting in a more accurate portrayal of both the dress and the person wearing it.
Virtual try-on technology aims to eliminate the guesswork when it comes to finding the right fit for customers of all body types. Various companies (Adobe, Amazon and Walmart) have launched their own tools, allowing customers to virtually try on all types of clothing, including dresses. However, with this new expansion, it seems that Google is looking to create a more advanced feature than its competitors.
