Research on automatic generation and style transfer of IP imagery based on deep learning

The growing demand for efficient and scalable solutions in creative industries has spurred significant advancements in artificial intelligence, particularly in the design and generation of IP imagery. Traditional methods for IP imagery design rely heavily on manual efforts, making the process time-consuming, resource-intensive, and difficult to scale. Recent developments in deep learning, such as generative adversarial networks (GANs) and diffusion models, have enabled automation and innovation in artistic workflows. However, these methods often fail to address the unique requirements of IP imagery, such as stylistic consistency, adaptability across diverse platforms, and alignment with branding objectives. This study proposes a novel deep learning framework for the automatic generation and style transfer of IP imagery. The framework integrates a transformer-based generation model with a hierarchical style transfer mechanism, enabling the production of high-quality, stylistically consistent outputs across various artistic domains. Hierarchical latent space encoding preserves global structural coherence and local stylistic detail, while adaptive instance normalization (AdaIN) facilitates seamless blending of content and style. Extensive experiments demonstrate the framework's superiority over baseline models, achieving lower Fréchet Inception Distance (FID), improved perceptual similarity, and enhanced style consistency. A user study further validates its practicality, highlighting its ability to meet diverse artistic and branding requirements. These results underscore the framework's potential in automating creative workflows and advancing IP design.