This paper presents a theoretical analysis of the impact of guidance on diffusion models, particularly in the context of Gaussian mixture models (GMMs). The study shows that incorporating guidance into diffusion models increases classification confidence and reduces distribution diversity, leading to a decrease in differential entropy of the output distribution. The analysis covers widely used sampling schemes such as DDPM and DDIM, and leverages comparison inequalities for differential equations and the Fokker-Planck equation to characterize the evolution of probability density functions. The results demonstrate that guidance enhances the alignment of generated samples with task-specific information, but can also lead to reduced diversity. The paper also highlights a complex interaction between guidance strength and the behavior of classification confidence, particularly in the case of three-component GMMs with aligned means. Theoretical guarantees are provided for the effects of guidance on classification confidence and distribution diversity, and the results are validated through numerical experiments. The study provides a deeper understanding of the theoretical underpinnings of diffusion guidance and its implications for generative modeling.This paper presents a theoretical analysis of the impact of guidance on diffusion models, particularly in the context of Gaussian mixture models (GMMs). The study shows that incorporating guidance into diffusion models increases classification confidence and reduces distribution diversity, leading to a decrease in differential entropy of the output distribution. The analysis covers widely used sampling schemes such as DDPM and DDIM, and leverages comparison inequalities for differential equations and the Fokker-Planck equation to characterize the evolution of probability density functions. The results demonstrate that guidance enhances the alignment of generated samples with task-specific information, but can also lead to reduced diversity. The paper also highlights a complex interaction between guidance strength and the behavior of classification confidence, particularly in the case of three-component GMMs with aligned means. Theoretical guarantees are provided for the effects of guidance on classification confidence and distribution diversity, and the results are validated through numerical experiments. The study provides a deeper understanding of the theoretical underpinnings of diffusion guidance and its implications for generative modeling.