Iterative Prompt Refinement for Text-to-Image Generation: A Multi-Metric Evaluation Study

Recent advancements in Generative Artificial Intelligence have significantly accelerated the development of text-to-image generation systems capable of producing high quality and visually realistic images from natural language prompts. Despite these advancements, the quality and accuracy of generated images remain highly dependent on the design of the input prompt, making prompt engineering a critical and challenging task. Users often need to modify prompts to achieve the desired output, which is both time consuming and computationally inefficient. This work proposes an efficient iterative prompt refinement framework based on BLIP (Bootstrapping Language-Image Pretraining) to improve the alignment between textual prompts and generated images. Unlike traditional approaches that rely on Large Language Models (LLMs) for prompt refinement, the proposed system utilizes a vision language model to analyze the generated image and extract a semantic caption. This caption is compared with the original or previously refined prompt to identify missing or mismatched details, enabling the system to progressively generate improved prompts. The iterative feedback loop allows the system to gradually enhance both semantic alignment and visual quality of the generated outputs. The system also achieves better efficiency compared to conventional methods by leveraging semantic feedback directly from generated images.