Research on image generation art creation assisted by deep learning and embedded systems

Abstract The use of state-of-the-art image generation models for artistic purposes on embedded devices remains challenging due to their high computational and power requirements. The goal of this research is to create a system for real-time image generation for artistic creation that integrates deep learning models with embedded hardware. The method suggested uses a diffusion-based image generation framework optimised through model pruning and TensorRT acceleration for inference on devices with limited resources. A varied dataset featuring a wide range of artistic styles was implemented to assess the visual quality and performance of the system. The experiment’s outcome shows that the system makes a Structural Similarity Index (SSIM) of 0.89 and a Fréchet Inception Distance (FID) of 16.7, while the average inference latency is 1.3 s per image and the power consumption is around 15 W. User ratings also point out that the users have a high level of satisfaction with both the quality and promptness of the images produced. These results demonstrate the feasibility of deploying diffusion-based image generation on embedded systems under strict latency and power constraints, providing an energy-efficient and portable solution for interactive and educational generative art applications.