Deep Diffusion Probabilistic Models (DDPMs) are a versatile class of generative models able to efficiently model complex data distributions and have quickly become challenging benchmarks for density estimation tasks. In this work, we present a neat solution to implement DDPMs for image generation using a compact TinyUNet architecture as well as diffusion-based sampling methods on probability modeling. The Fashion MNIST dataset is the canonical training data used for our model, which leverages a back propagation-based loss minimization algorithm implemented using TensorFlow to iteratively optimize encoder/decoder parameters so as to minimize Mean Squared Error between predicted noise and ground truth noise. We comprehensively evaluate and compare our method to demonstrate its effectiveness for synthesis of high-quality images with different visual attributes. In addition, we offer insights into how the DDPM framework operates, demonstrating model architecture interpretability and scalability. Our work is a good snapshot in the much larger context of generative modeling techniques and will provide clues for this type to scale further.