Diffusion Models For Image Generation: A Review Of Denoising Probabilistic Models, Latent Diffusion, Guidance And Controllable Synthesis

Diffusion models are generative models that create high-quality, realistic images by learning to reverse a gradual noising process. This paper reviews the working principles, architectures and applications of diffusion models for image generation. It explains the forward process, which adds Gaussian noise to training images over many steps, and the reverse process, in which a neural network, usually a U-Net, learns to predict and remove noise. The review traces the field from its thermodynamic origins to denoising diffusion probabilistic models (DDPM), the score-based stochastic differential equation view, and faster samplers such as DDIM. It examines classifier and classifier-free guidance, latent diffusion models that operate in a compressed latent space for efficiency, text-to-image systems conditioned on language embeddings, and spatial control methods that add edges, poses or depth as conditions. Applications in digital art, image editing, restoration, medical imaging and animation are discussed. A five-stage architecture is presented, along with evaluation metrics, research gaps, objectives, a comparison with GANs and VAEs and challenges such as slow sampling, computational cost, bias, copyright and misuse for deepfakes. The review concludes that diffusion models offer an excellent balance of quality and diversity, with efficiency and responsible use as the main open problems.