Text-Guided Image Generation Method Based on Cross-Modal Fashion Style Learning

In the interdisciplinary field spanning computer vision and art, text-guided image generation techniques have garnered significant attention, particularly in the realms of virtual try-ons and fashion design. Current approaches to text-guided image generation primarily utilize latent space encoding and alignment of text-image feature spaces, which prove effective in general scenarios. However, when tasked with generating images possessing specific fashion attributes, these methods often result in semantic inconsistencies and missing details in the generated clothing images. To address this challenge, this work presents an effective text-guided image generation method based on cross-modal fashion style learning. First, the fashion text encoding module is constructed to represent the input text description as global-local text descriptors to obtain fashion features. Then, a clothing image style representation module is designed to extract coarse-grained, fine-grained, and hierarchical latent features from the input texture patterns and clothing images. Finally, a cross-modal fashion style learning block, integrating the fashion mapping network and the generative adversarial network, is utilized to reinforce semantic consistency and yield clothing images with intricate details. Extensive experiments conducted on the Viton-HD dataset validate that the proposed method significantly outperforms other state-of-the-art approaches, achieving an initial IS score of 5.22 and a Fréchet Inception Distance (FID) of 16.35. Ablation studies and comparative analysis further demonstrate the potential application of our method, specifically in improving the semantic consistency between fashion texts and generated images, as well as enhancing fashion details in the generated images.