Abstract
In this paper, we share our experience in using synthetic image generation through text-to-image AI models with the aim of improving the performance of image classifiers models.The procedure for generating high-quality synthetic images with text-to-image AI models is described, utilizing generic prompts based on proximity in hypernymys to generate 200 images of samples (lions, cats, and monkeys). Two image classifiers were built using the synthetic images alongside with classical image augmentation techniques. Each classifier was evaluated using three different datasets. Stable Diffusion 2.1 was employed as the image generator, and the results demonstrate that our proposed technique of generating synthetic images can indeed enhance image classifier accuracy. The results rely on the proficiency of the AI and the appropriateness of utilized prompts. We think our results can be improved by advances on the Latent Diffusion models.