Controllable image generation and editing

Abstract

This paper presents a comprehensive exploration of text-to-image generation, examin- ing the current state-of-the-art models, available fine-tuning techniques using LoRAs, and the limitations that come with them. In addition to this survey, we propose a method that enhances user control over image outputs, enabling more targeted and cus- tomizable generations. Our method demonstrated strong performance across various styles and subjects. We also explored how a distinction between styles highlights the importance of adapting generation approaches based on stylistic intent. We also outline key strategies and fine-tuning practices that can significantly improve generation qual- ity, regardless of the underlying model or method. These insights aim to equip users with both the understanding and practical tools needed to optimize their text-to-image workflows.