Generating High-Resolution Fashion Model Images Wearing Custom Outfits

Gökhan Yildirim,Nikolay Jetchev,Roland Vollgraf,Urs Bergmann

DOI: https://doi.org/10.48550/arXiv.1908.08847

2019-08-23

Abstract:Visualizing an outfit is an essential part of shopping for clothes. Due to the combinatorial aspect of combining fashion articles, the available images are limited to a pre-determined set of outfits. In this paper, we broaden these visualizations by generating high-resolution images of fashion models wearing a custom outfit under an input body pose. We show that our approach can not only transfer the style and the pose of one generated outfit to another, but also create realistic images of human bodies and garments.

Computer Vision and Pattern Recognition,Machine Learning,Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: the need for visualizing customized clothing combinations (outfit) on fashion e - commerce platforms. Specifically, due to the complexity of combining different clothing items and the limited existing image resources, only preset combination schemes can be shown. To improve this situation, the author proposes to generate high - resolution fashion model images. These images can show user - customized clothing combinations and can be adjusted according to the input body postures.

### Main problems and solutions

1. **Existing limitations**:
- Existing fashion e - commerce platforms can only provide a limited number of preset combination pictures.
- Previous studies focused on replacing a certain piece of clothing in existing model pictures [5, 2] or generating low - resolution images from scratch [8].

2. **Research objectives**:
- **Generate high - resolution images**: Through generative adversarial networks (GANs), especially the improved Style GAN, generate high - resolution fashion model images to show customized clothing combinations.
- **Support different body postures**: Not only be able to generate static model images, but also be able to dynamically adjust the images according to the input body postures.

3. **Specific methods**:
- **Unconditional Style GAN**: Train a standard Style GAN model to generate high - quality model images, and achieve style/color and pose transfer by exchanging style vectors in specific layers.
- **Conditional Style GAN**: Introduce an embedding network, and take clothing items and human postures as conditional inputs, so as to generate model images with specific combinations and postures.

### Experimental results
- **Unconditional Style GAN**: The generated images have a high sense of reality at a resolution of 1024×768, including details of clothing and human body parts.
- **Conditional Style GAN**: It can generate realistic model images according to the input clothing items and postures, and can handle models of different body types.

### Conclusions
The author shows two methods for generating high - resolution fashion model images, one is the unconditional Style GAN, and the other is the conditional Style GAN. The former can achieve style and pose transfer, and the latter can generate customized model images according to the input clothing and postures. Future work will focus on improving the image quality and consistency of the conditional model, especially in the case of dealing with complex textures and texts.

Through these methods, this research is expected to significantly improve the user experience of fashion e - commerce platforms, enabling users to more intuitively see the effects of clothing combinations they are interested in.