Text-to-image generation has grown into an important application of generative AI. Among the many models, diffusion-based architectures such as Stable Diffusion and Dreamlike Diffusion have become popular. They can create high-quality images with realistic and stylistically rich visuals using natural language prompts. In this work, we assess and contrast several well-known diffusion-based models. We create images and evaluate their quality using perceptual and semantic evaluation metrics, such as CLIP score, LPIPS, FID, and Inception Score, using a uniform prompt describing a forest scene. Our analysis shows that while Dreamlike Diffusion excels in artistic coherence and perceptual detail, models like Stable Diffusion XL offer balanced trade-offs between realism and prompt alignment. For instance, cases involving the production of artistic images, the results provide valuable information regarding model selection. This interpretational study can act as a guide for researchers and practitioners who want to compare generative models with standard metrics highlighting fidelity.
Paper
Full text
Performance Analysis of Open-Source Diffusion Models for Artistic Image Generation
Semantic Scholar · 2025
Abstract
Text-to-image generation has grown into an important application of generative AI. Among the many models, diffusion-based architectures such as Stable Diffusion and Dreamlike Diffusion have become popular. They can create high-quality images with realistic and stylistically rich visuals using natural language prompts. In this work, we assess and contrast several well-known diffusion-based models. We create images and evaluate their quality using perceptual and semantic evaluation metrics, such as CLIP score, LPIPS, FID, and Inception Score, using a uniform prompt describing a forest scene. Our analysis shows that while Dreamlike Diffusion excels in artistic coherence and perceptual detail, models like Stable Diffusion XL offer balanced trade-offs between realism and prompt alignment. For instance, cases involving the production of artistic images, the results provide valuable information regarding model selection. This interpretational study can act as a guide for researchers and practitioners who want to compare generative models with standard metrics highlighting fidelity.