TryOnDiffusion_云技派,专注ai人工智能! 207

TryOnDiffusion is a diffusion architecture based on two UNets, designed to generate visual renderings of a garment on the body of the person shown in the input image. The method is capable of preserving garment details and adapting to significant changes in body pose and shape. TryOnDiffusion achieves state-of-the-art performance both qualitatively and quantitatively, making it a highly effective virtual try-on technology.

TryOnDiffusion can combine two images to change the outfit of the person in a photo. One image shows a person, while the other shows a garment worn by someone else—it then generates an image showing what the garment might look like on the input person. Previously, such models often suffered from a lack of garment details. Traditional composition methods usually only worked with two images whose poses and shapes were fairly similar; otherwise, the newly composed image would show severe garment distortion. TryOnDiffusion, by contrast, leverages a diffusion architecture called Parallel UNet, enabling the composed image to retain garment details while adapting to changes in the person's pose and body shape.

Features of TryOnDiffusion:

• The model uses diffusion techniques to show how clothing looks on people of different body types and poses.

• The new ai model is capable of generating high-quality, realistic try-on images.

• The model uses image diffusion and cross-attention techniques rather than text input.

• To improve accuracy, the model is trained using Google's Shopping Graph.

• The virtual fitting room function currently supports women's tops from brands such as Anthropologie, LOFT, H&M, and Everlane.

• The feature will gradually be expanded to more brands.

• It is currently not open source.