SP3F-GAN: GENERATING SEAMLESS TEXTURE MAPS FOR FASHION

Creating seamless textures is crucial to attain realistic 3D virtual objects and environments. This is because when stochastic textures are tiled in a straightforward manner, they produce cluttered visuals with noticeable seams that lack authenticity. This paper proposes GAN-based method for automatic seamless texture synthesis by designing three main components in generator block: (i) the texture style encoder, (ii) residual tiling blocks and (iii) the BRDFs tileable decoder in the generator of an adversarial expansion network, resulting in a continuous texture output at the seam intersection area. In addition, considering the different properties of virtual environment rendering materials, the spatially varying BRDFs (albedo, normal, roughness and displacement map) are all designed in the proposed model to handle multi-layer texture representation. Furthermore, the proposed method for generating seamless multi-layer texture maps that incorporates different loss functions, allowing us to control both the geometric shape and visual fidelity of the synthesized textures. Qualitative and quantitative experiments on the describable textures dataset (DTD) show that the generated texture maps are not only seamlessly tiled but also exhibit superior visual quality in preserving details compared to previous deep texture synthesis methods.