âControlled GAN-Based Creature Synthesis via a Challenging Game Art DatasetâAddressing the Noise-Latent Trade-Offâ, 2021-08-19 (; backlinks; similar)â :
The StyleGAN-2 network supports powerful methods for creating and editing art, encompassing generating random images, finding images âlikeâ some query, and modifying content or style. Additionally, recent advancements have enabled training with small datasets. We apply these techniques to synthesize card art, focusing on a novel Yu-Gi-Oh dataset.
While noise inputs to StyleGAN-2 are crucial for effective synthesis, we discovered that, in the context of small datasets, coarse-scale noise conflicts with latent variables as both influence long-scale image effects. This issue manifested as over-aggressive variation in art following changes in noise and diminished content control when editing latent variables. To address these challenges, we trained a modified StyleGAN-2 model in which coarse-scale noise is suppressed, thereby eliminating these undesirable effects.
Through this process, we achieved a superior FID, indicating improved image quality. Moreover, alterations in noise now facilitate local exploration of style rather than causing erratic changes, and identity control through latent variable adjustments is significantly enhanced.
Our results and analyses serve as foundational steps towards the development of a GAN-assisted art synthesis tool. Such a tool holds immense potential for digital artists across various skill levels, offering new horizons in artistic ideation within film, games, and other creative industries.