GSCNet: A transformer-based granular style control network for artifact-free image style transfer
Zheng Gao, Zhicheng Bao, Haoran Fan, Xiaoyu Li, Qipei Nong, Jiaojiao Jiang · Journal of Visual Communication and Image Representation · 2026
Transformer-based image style transfer models capture long-range dependencies but often produce visual artifacts due to insufficient granularity control. We propose GSCNet, a granular style control network that suppresses artifacts from three complementary perspectives. For structural artifacts from limited receptive fields in window-based attention, we design a Dual-grained Attention Module (DAM) that uses global context to guide local feature interactions. For pixel-level artifacts such as uneven stylization from abrupt value shifts, a Feature Stabilization Module (FSM) normalizes outlier activations using neighborhood information. For fragmented detail artifacts caused by overfitting, a Detail Engraver Adapter with LoRA enhances fine-grained perception while maintaining coherent style fusion. Extensive experiments demonstrate that GSCNet leads on four of the six evaluation metrics over the strongest baseline (S2WAT), with up to 41% lower feature-level identity loss (Identity Loss 2); as a rough aggregate, this corresponds to a mean improvement of 12% across the six metrics. The gains are concentrated in content preservation, identity consistency, structural similarity, and reconstruction fidelity, while the higher style loss reflects a deliberate content–style trade-off.