Research on artificial intelligence: driven generative design for digital art and creative visualization

Abstract

Generative design achieved through Artificial Intelligence has become an innovative practice in digital art and creative visualization, as it allows creating high-quality artistic content automatically and with minimal human intervention. In this paper, the authors will come up with an Improved Multi-Scale ArtFusionNet-based Diffusion Model (MS-AFN-Diff) that will be used to improve the quality, variety, and semantic consistency of created artworks. The model incorporates the use of multi-scale feature extraction to capture the global structure and fine-grained artistic detail, and an ArtFusionNet module to blend both style and content. One more optimization measure is included to optimize the generative procedure and enhance the convergence efficiency. The StyleBreeder Dataset (2024) consists of a huge amount of different artistic imagery with various styles, textures, and compositions, which is used as the experimental evaluation. The dataset is systematically preprocessed and resized to a common resolution (256 × 256), normalized, and augmented (rotation, flipping, and jittering color), and with noise to improve generalization. To guarantee sound model testing, the data set is divided into training (70%), validation (15%), and testing (15%). The evaluation of quantitative performance is performed based on traditional metrics, including Fréchet Inception Distance, Inception Score, CLIP Score, and Structural Similarity Index Measure (SSIM). The proposed MS-AFN-Diff model reaches the value of 7.8, which is a significant improvement when compared to baseline models like Stable Diffusion(12.5) and GAN-based models (greater than 15). Moreover, it has a better Inception Score of 24.3, which means an increased diversity and realism. The CLIP score of 0.36 shows better semantic matching between the generated images and textual descriptions, whereas the SSIM score of 0.87 shows better preservation of the structure. All in all, the suggested model offers a trade-off between computational efficiency and image quality, which is rather reasonable and makes it very applicable in various applications in digital art, animation, and creative design systems.