Article: Oil painting style transfer with staged spatial-channel attention in encoder and decoder Journal: International Journal of Reasoning-based Intelligent Systems (IJRIS) 2026 Vol.18 No.17 pp.29 - 47 Abstract: Existing attention-based generative adversarial networks for oil painting style transfer typically employ a single type of attention or stack them in parallel, failing to simultaneously preserve global composition and render fine brushstroke textures. This paper proposes a dual-path attention generative adversarial networks that introduces a staged attention fusion strategy: a spatial attention module is embedded after the encoder to model long-range dependencies and preserve global structure, while a channel attention module is placed before the decoder to recalibrate texture-sensitive feature responses. This staged design explicitly addresses the gap that existing methods do not differentiate attention placement according to the distinct requirements of encoding (structure preservation) and decoding (texture generation). In addition, an attention-guided content preservation loss is proposed to supervise attention distribution using divergence, preventing semantic distortion. On the Monet2Photo dataset, dual-path attention generative adversarial networks reduces Fréchet inception distance from 78.3 to 68.4, a 12.7% improvement. In a user study, 71% of generated images were misclassified as human-painted artworks. These results demonstrate that staged attention with explicit supervision bridges the structural-textural gap in oil painting style transfer. Inderscience Publishers - linking academia, business and industry through research

Title: Oil painting style transfer with staged spatial-channel attention in encoder and decoder

Authors: Jianren Qi; Li Gao; Congcong He

Addresses: Art Design Department, Shandong Vocational College of Special Education, Jinan, 250306, China ' Art Design Department, Shandong Vocational College of Special Education, Jinan, 250306, China ' Art Design Department, Shandong Vocational College of Special Education, Jinan, 250306, China

Abstract: Existing attention-based generative adversarial networks for oil painting style transfer typically employ a single type of attention or stack them in parallel, failing to simultaneously preserve global composition and render fine brushstroke textures. This paper proposes a dual-path attention generative adversarial networks that introduces a staged attention fusion strategy: a spatial attention module is embedded after the encoder to model long-range dependencies and preserve global structure, while a channel attention module is placed before the decoder to recalibrate texture-sensitive feature responses. This staged design explicitly addresses the gap that existing methods do not differentiate attention placement according to the distinct requirements of encoding (structure preservation) and decoding (texture generation). In addition, an attention-guided content preservation loss is proposed to supervise attention distribution using divergence, preventing semantic distortion. On the Monet2Photo dataset, dual-path attention generative adversarial networks reduces Fréchet inception distance from 78.3 to 68.4, a 12.7% improvement. In a user study, 71% of generated images were misclassified as human-painted artworks. These results demonstrate that staged attention with explicit supervision bridges the structural-textural gap in oil painting style transfer.

Keywords: generative adversarial network; GAN; attention mechanism; oil painting style transfer; image generation.

DOI: 10.1504/IJRIS.2026.154539

International Journal of Reasoning-based Intelligent Systems, 2026 Vol.18 No.17, pp.29 - 47

Received: 14 Mar 2026
Accepted: 11 Apr 2026
Published online: 02 Jul 2026 *

Title: Oil painting style transfer with staged spatial-channel attention in encoder and decoder

Keep up-to-date