Análise da Geração de Imagens a Partir da Descrição em Texto de Cenas Utilizando Stable Diffusion

Abstract

The field of synthetic image generation, using Artificial Intelligence (AI), has made signi f icant advances in recent years, with notable contributions from models such as DALL-E, MidJourney, Stable Diffusion, and others. This work proposes to explore the current state of the art in image generation from natural language descriptions, emphasizing the techniques employed and approaches adopted in the area. The Stable Diffusion, a po pular Latent Diffusion Model and reference in the Open Source scope, which stands out for performing synthetic image generation in a latent space, with low inference time and computational cost, is the main focus of this study. The conducted experiments presented variations of items to be generated, quantity, and color, initially with elementary textual descriptions, such as “dog”and “cat”, and advancing to more detailed scenarios, such as “ten red dogs and ten blue cats”. This variation allowed for a deep qualitative analysis of the impact of scene descriptions on the results obtained by the Stable Diffusion model. The study identified significant challenges in the area, especially in optimizing scene des criptions for generative models. The discovery of best practices for formulating prompts is an evolving process and is crucial to achieving the expected results in the generated images. Another relevant issue is the model’s limitation in producing realistic images faithful to the requested details when presented with prompts that contain a high number of objects to be represented. The research concludes that the appropriate selection of text descriptions is essential to guide the image generation process and achieve the desired re sults. However, despite significant advances, the field still demands additional research to overcome these challenges and improve the quality of the generated images.