Rotate your phone to navigate through the site properly

TextAIles: Color Manipulations in the Latent Space of StyleGAN3

Ludovica Schaerf will present the Paper “TextAIles: Color Manipulations in the Latent Space of StyleGAN3”  tomorrow Tuesday January 23, 2024, at the Conference  Computer Vision and Image Analysis of Art, in San Francisco. The paper is part of the homonymous session taking place 3:30 PM – 5:30 PM / Room: Harbour B, and chaired by David G. Stork.


TextAIles: Color Manipulations in the Latent Space of StyleGAN3 (CVAA-175)

Ludovica Schaerf, Dario Negueruela del Castillo, Max Planck Society, and Leonardo Impett, University of Cambridge

Abstract: Artificial Intelligence is increasingly being consumed in artistic and design practices. However, our understanding of AI models and their inner workings in relation to fundamental aspects of those practices is largely unknown and underexplored. Comprehending and enhancing crucial abstractions in the models – the controllable aspects of the world they create – becomes essential.
The contribution of this work is twofold: first, we utilize state-of-the-art disentanglement methods for color editing of textile patterns using StyleGAN3’s latent space; and second, we use these disentangled abstract representations to investigate StyleGAN3’s internal representations of color systems, hues, and palettes. To the best of our knowledge, no disentanglement experiments have been carried out in the context of color editing and textiles.
The result is an efficient, training-free, and adaptive color editing tool with a vast range of applications, from intelligent colorway creation for textile patterns to colorizing tools from black-and-white images or faded materials. Furthermore, the tool sheds light on the cultural paradigms of colored patterns as mirrored in the models; on the latent similarity of the colors in this space; and on their combinatorial properties.