SKTNet: a semantically guided attribute disentanglement network for fashion sketch editing

Fashion sketch editing is intended to modify specific attributes of a sketch while preserving its original integrity, thus facilitating the rapid transformation of designers’ concepts into tangible designs. When fashion sketches are edited, it is crucial to precisely control the style of different parts and ensure that line connections and transitions are smooth and natural. The presence of complex and diverse semantic attributes in fashion sketches poses a challenge to focus existing attribute editing efforts on specific regions. To overcome this limitation, a semantically guided fashion sketch disentanglement model is proposed. First, the complete fashion sketch undergoes semantic segmentation into sleeves and torso using a semantic segmentation network. Thereafter, the latent space of the network is decomposed into semantic parts based on semantic segmentation to prevent editing operations from having unnecessary or unintended effects on non-targeted regions of the fashion sketch. Subsequently, a VGG-structured encoder and StyleGAN2 decoder are trained to obtain the latent vectors of both the complete and segmented sketches. More concise and explanatory features are then extracted in the latent space through sparse principal component analysis. Finally, perturbations along the principal directions are applied to explore variations in attributes related to sleeves and torso in fashion sketches. Extensive qualitative and quantitative experiments on the VITON-HD and Dress-Code datasets demonstrate that our model exhibits outstanding disentanglement ability and produces excellent editing effects in the target attribute regions while keeping the non-target regions virtually unaltered. Furthermore, the attribute disentanglement accuracy is significantly higher than that of other methods.