Recent advancements in deep generative models have spearheaded the development of novel techniques that redefine the standards in various tasks of machine learning. Among these, diffusion models have emerged as a class of generative models that facilitate gradual transformation from noise to data distribution through a Markov chain process. Originally conceived for generating high-quality images, diffusion models have shown promising results in generation tasks. This study extends the application of diffusion models beyond data generation, primarily focusing on the task of classification. We propose a new paradigm wherein diffusion models are trained in concert with a discriminative classifier, harnessing their generative capabilities to enrich the classifier's decision boundary. By reversing the diffusion process, the model generates intermediate representations that carry salient features essential for class discrimination. Our approach entails a stepwise refinement of the data manifold by traversing backwards through the diffusion process. By doing so, it provides the classifier with a dynamically enriched representation of the input, offering enhanced robustness and improved generalization capabilities. We evaluate the performance of diffusion models in a classification context using several benchmark datasets, showcasing their ability to improve classification accuracy, especially in scenarios of limited labelled data and class imbalance. Our results indicate that diffusion models can significantly boost the performance of classifiers through unsupervised learning from noise distribution to the formation of class-specific features. Furthermore, we explore the impact of the diffusion model's hyperparameters on classification outcomes and provide insight into the relationship between the quality of generated samples and classification performance. This study opens a new avenue in the application of diffusion models and sets the stage for further exploration of such models for tasks traditionally dominated by discriminative approaches. Through experimental trials and comprehensive analysis, we demonstrate that the intricate entwinement between generative and discriminative modelling can yield state-of-the-art results, ultimately advancing the frontiers of classification tasks in machine learning.