Enhancing Text-to-Image Model Evaluation: SVCS and UCICM

As the world gradually embraces the integration of AI into various applications, there is a growing need for methods to enhance and evaluate AI-based technologies. One such a remarkable innovation is text to image models, which are currently revolutionizing the content generation, computer vision, design industries etc. Among the existing deep learning techniques utilized for text to image synthesis, stable diffusion model set apart due to its ability to generate high quality images. It transforms images into their latent space representation, mitigating computational memory and time requirements while minimizing instability and complexity. In this paper, we propose two evaluation metrics Semantic Visual Consistency Score (SVCS) and User-Centric Image Coherence Metric (UCICM) to analyze the coherence of the image with respect to the user prompt. SVCS assesses the model's capability to interpret human language, while UCICM is tailored to enhance application-specific image generation through personalization. This paper primarily focuses on the processes entailed in a transfer learning-based stable diffusion model and subsequently evaluates its performance using the newly proposed metrics.