Large-size multivariate scale-free texture generation from small-size learning sets with self-attention Wasserstein-GAN

This work combines convolutional Wasserstein-GAN architectures with self-attention to generate multivariate scale-free textures at sizes larger than those in the training set. It evaluates how well the generated textures preserve multiscale and cross-component statistics, and compares self-attention in the generator, critic, and both networks. Across ten independent trainings, the proposed architecture generates textures up to eight times larger in each dimension than the originals, with the most consistent results when self-attention is placed in the critic.

Recommended citation: Jhonatan Ancco Cerda, Kinan Abbas, Stéphane G. Roux, and Patrice Abry. (2026). "Large-size multivariate scale-free texture generation from small-size learning sets with self-attention Wasserstein-GAN." Submitted to ICASSP 2027.