Coarse-to-fine clothing image generation with progressively constructed conditional GAN

Abstract

Clothing image generation is a task of generating clothing product images based on input fashion images of people dressed. A result of existing GAN based methods often contains visual artifact with the global consistency issue, due to training instability of GAN. To solve this issue, we split the difficult single image generation process into relatively easy multi-stages for image generation process. We apply a coarse-to-fine strategy on an image-conditional image generation model, with a multi-stage network training method, called rough-to-detail training. We also design our generator architecture appropriate for rough-to-detail training, by progressively configuring a target image of each stage through adding a decoder block. Via the coarse-to-fine process, our model can generate from small size images with rough structures to large size images with details. To validate our proposed model, we perform extensive evaluations on the LookBook dataset. Compared to other methods, our model can create visually pleasing 256 × 256 clothing images while keeping the global structure and containing details of target images.