Comparative Analysis of Text-to-Image Generation Models

Text-to-image generation is one of the emerging fields in the generative artificial intelligence (GAI). Most of the models focus on the realism and the visual fidelity of the image but fail to capture semantic meaning behind it. So, the major challenge for the models is to ensure semantic coherence and contextual relevance between the text and the generated image. This work utilizes two state-of-the-art text-to-image generative models, Stable Diffusion & Kandinsky for comparative analysis. Both of these models are fine-tuned on the MJ-Bench alignment dataset using LoRA. Different quantitative measures, including CLIP score, SSIM, and LPIPS, and qualitative evaluation through visual comparison and human evaluation are used. The results of the experiment provide an insight into the strengths and limitations of the different models, and also show the impact of fine-tuning on the quality and fidelity of the image, providing a better understanding of how each model behaves in text-to-image generation.