Detailed text-to-image generation

Abstract

Logo generation models have been assisting designers in creating aesthetically pleasing domain-specific logos. However, these models often lack the ability to produce unique designs as they typically rely on predefined templates. In contrast, popular image generation models like DALL-E and Stable Diffusion demonstrate exceptional capabilities in generating high quality images but struggle when it comes to creating coherent visual text and unique logos. Logo image-caption datasets have also been scarce, which are crucial for finetuning image generation models to improve their visual text generation capabilities. This report details the collection and processing of logo images and their captions to develop a high quality, well-annotated dataset. To evaluate the quality and effectiveness of the dataset, it is used to finetune text-to-image generation models, to assess the improvement of the finetuned models' text generation ability in images.