A disentangled CNN–transformer framework for big data-driven intelligent brand visual generation
Year: 2026
Type: article
Abstract: Big data has driven the rapid growth of brand visual design demand, yet existing intelligent generation methods suffer from detail distortion and poor cross-cultural adaptability. This study proposes a style-disentangled Convolutional Neural Network (CNN)-Transformer framework for intelligent brand visual generation. A dual-branch architecture extracts local details via CNN and global style via Transformer, with an embedding mixer enabling balanced feature fusion. A regional gating and style calibration module further enhances structural fidelity. Evaluated on a self-constructed 10-brand multimodal dataset, the model outperforms state-of-the-art methods on Learned Perceptual Image Patch Similarity (LPIPS) (0.186) and Structural Similarity Index Measure (SSIM) (0.924), with style consistency exceeding 85% across cross-cultural scenarios. Ablation studies have shown that, compared with the full model, only the CNN branch increases LPIPS by 54.3%, confirming the contribution of key modules. This work provides an effective solution for big data-driven intelligent brand design and extends computational creativity applications.
Source: Scientific Reports
Authors: Yanan Jiang, Yongxiao Liu
Institution: Baoshan University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7223173194