AttCloth-GAN: an improved attention-based conditional generative adversarial network for extracting clothing from specified body part categories in full-body model images

Purpose People are often drawn to models' attractive clothing, which sparks their desire to search for and purchase such items. However, consumer preferences vary – some favor full outfits, while others focus only on lower garments. Tiled clothing images greatly boost search success rates. Thus, we propose an improved attention-based conditional GAN to generate such tiled images, which can produce clear, accurate, category-compliant results from models' full-body images. Design/methodology/approach We propose a two-stage framework, where the first stage generates coarse images and the second stage generates fine images: Stage 1 employs CBAM attention to enhance skip connections in U-Net, suppressing irrelevant information like human bodies while sharing low-level features. Labels are concatenated at the bottleneck layer to specify the generated clothing category. We modified the original residual blocks by integrating CBAM attention into them and applying this to Stage 2 of our model, mitigating gradient vanishing while further suppressing irrelevant information. The improved network, termed AttCloth-GAN, can generate clothing images of specified categories. Findings Experimental results demonstrate that our method achieves optimal performance while generating images that align with category labels. On the dresscode dataset, our model improves upon pix2pixHD by 3.0% in SSIM and reduces FID by 10.0%. Originality/value Our proposed method can convert full-body model images into tiled clothing images of specified categories, meeting diverse user needs. Users can extract clothing from their preferred body parts for retrieval, significantly enhancing search accuracy and effectively promoting the development of online clothing shopping.