Google has recently launched a new feature in their shopping platform that allows users to try on dresses virtually. This innovative tool is powered by a generative AI technology developed specifically for virtual fittings, utilizing a diffusion-based technique to generate lifelike images of dresses on models.

During the development process, the team encountered challenges unique to dresses. Unlike simple tops, dresses are more intricate in their design, with variations in draping, silhouette, length, and shape. Trying to replicate these details in a virtual space posed a significant challenge, as enlarging the image did not necessarily enhance the clarity of the finer elements, such as floral prints or ruffled collars.

To address this issue, a “progressive training strategy” was implemented, allowing the AI model to progressively enhance the resolution of the images during the diffusion process. This approach ensures that even the smallest details, such as pleats and prints, are accurately represented in the virtual try-on experience.

Another obstacle faced was the fact that dresses cover a larger portion of the body compared to tops, making it challenging to replace the dress without affecting the person’s features or body details. To overcome this, a technique called the VTO-UNet Diffusion Transformer (VTO-UDiT) was developed to isolate and preserve the essential features of the person wearing the dress.

By incorporating the VTO-UDiT technique, the model can focus on maintaining the individual’s identity while refining the representation of the dress. This process not only enhances the accuracy of showcasing the garment but also ensures that the wearer’s face and body are accurately depicted in the virtual try-on experience.

This groundbreaking technology aims to revolutionize the online shopping experience, providing users with a more realistic and personalized way to preview dresses before making a purchase. To learn more about this innovative feature, visit the official Google blog here.