توليد صور المشروطة بالنص باستخدام نماذج الانتشار

Text - Conditioned Image Generation using Diffusion Models

Abstract

Text-guided synthesis of images has made a giant leap toward becoming a mainstream phenomenon. With text-to-image generation systems, anybody can create digital images and artwork. This provokes the question of whether text-to-image generation is creative.
The generative systems have contributed much to the development of artificial intelligence (AI) generating rather realistic images from the text. Text-to-image generation systems have been used in various forms and areas in scope including, but not limited to, artworks and designs, data sampling, and entertainment. Many studies have been conducted on generating images from text and many AI techniques have been proposed. However, some critical issues have yet to be solved, especially with regard to the time consumption and the training time. Therefore, the proposed study utilized the Stable Diffusion Model (SDM) to conduct iterative feedback (if the metrics of the evaluation namely Inception Score (IS) and Fréchet inception distance (FID) do not improve then the hyper-parameters are tuned and the model is trained again). In this study, the fine-tuning of the SDM results in a considerable improvement in generating images that are more akin to reality. As well, there are trade-offs between image quality and flexibility in performance metrics. The fine-tuning process gradually improves the model’s global ability to generate better and more diverse digital imagery. The fine-tuned model has a lower FID score (248.748256), suggesting a higher likelihood of attaining higher image distribution similarity to the targeted dataset. Sparingly, the results of the improved model denoted a lower FID score (212.52) when contrasted with the base model (251.22), pointing out that the generated images from the fine-tuned model were more intimate to the target distribution in the synthetic dataset.