Virtual Try-on based on Generative Adversarial Networks

---

Abstract

In this work, we transform the virtual try-on mission into a fashion image completion task. That is, given a portrait missing a piece of clothing, we want to generate a complete fashion image with high realism, diversity and compatibility. By proposing a two-stage fashion image completion network, we divide the generation process into two sub-processes: shape generation and texture synthesis, which realize hierarchical control of shape and texture, and improve the realism of the generated image. In each sub-process, we use the idea of Variational Autoencoder: map the control information to the latent space and then utilize the random sampling to realize the variety of generated images. At the same time, we introduce two interrelated encoders to explicitly control the compatibility among clothing items. In addition, we use methods such as generative adversarial networks, attention mechanism, and spectral normalization to further improve the quality of generated images and stabilize the training process. We test our model on DeepFashion dataset and compare our model with other similar work, the result of which confirm that our model has excellent performance in the fashion image completion task.

Bachelor Thesis: [PDF]       Code: [GitHub]

Experiment Results

Comparison of different image generation methods and ablation test.