Adversarial Training Inspired Self-attention Flow for Universal Image Style Transfer

Flow-based model receives more and more attention and has been recently applied to image style transfer. While these methods can achieve splendid performance, there remains a problem that the stacked convolutions are inefficient and cannot focus on valuable features. Starting with training an adversarial robust model, we find that no matter in the perceptual loss network or the transfer model, robust features are beneficial for performing better universal style transfer (UST) results. Based on this initial conclusion, we improve the current Glow model by applying self-attention mechanism with three different blocks using ViT, non-local and involution, respectively. Designed feature extraction blocks can capture more valuable deep features with fewer parameters, making Glow more effective and efficient in UST. Our improved Glow can generate artistic images that look nicer and more stable. Both visual results and quantitative metrics are compared to prove that our improvement makes Glow more suitable for UST.