Diffusion models are a generative AI method that learns to create new data by reversing a gradual noising process, starting from random noise and iteratively denoising it into a realistic image, sound, video or other structured output. Instead of predicting a category or a value, they generate entirely new examples that resemble the data they learned from. This approach has become one of the dominant architectures in generative AI, first powering image generation systems and then expanding into video generation, audio synthesis, scientific research and other domains. During training, an image, audio clip or other data sample is progressively corrupted until it resembles random noise, and the model learns to predict how each stage of corruption can be reversed. Once training is complete, it can begin with random noise and iteratively reconstruct a new sample. The same architecture has been adapted for video, speech, music, molecular design, protein generation and other domains. Diffusion models were not the first successful generative models, as variational autoencoders (VAEs) and generative adversarial networks (GANs) had already demonstrated that neural networks could produce convincing synthetic images, faces and artwork, though GANs were notoriously difficult to train, often suffering from unstable optimization or limited output diversity.