Definition
A diffusion model is a generative model that creates new data by learning to reverse a gradual noise-adding process. During training it observes how clean data turns into noise; at generation time it runs that process backwards, starting from pure noise and denoising it step by step until a coherent image, audio clip, or video emerges. Diffusion models are the technology behind systems like Stable Diffusion, DALL-E, and many modern audio and video generators.
The forward (noise-adding) process
The forward process is a fixed Markov chain that takes a real data sample and adds a small amount of Gaussian noise at each of many timesteps. After enough steps the original signal is completely destroyed and the result is indistinguishable from random noise. Crucially, this process has no learnable parameters — it is a known schedule — which means you can mathematically express the noisy version of a sample at any timestep directly, making training efficient.
The reverse (denoising) process
The reverse process is what the network actually learns. Given a noisy input and the timestep it came from, the model is trained to predict the noise that was added (or equivalently, the slightly-cleaner version of the data). Because each forward step added only a little noise, each reverse step only has to remove a little — a much easier task than generating a full image in one shot. By chaining these denoising steps from pure noise back to a clean sample, the model synthesises new data that matches the training distribution.
Why the step-by-step approach works
Generating high-quality data in a single pass is extremely hard. Diffusion models sidestep this by decomposing generation into hundreds of small, tractable denoising steps. Each step is a simple regression problem, so the training objective is stable and well-behaved — unlike the adversarial min-max game in GANs. The trade-off is sampling speed: producing one sample requires many model evaluations, which is why techniques such as DDIM sampling and distillation exist to cut the number of steps needed.
Why it matters
Diffusion models now dominate generative imaging and are spreading into audio, video, and even molecule design. They can be conditioned on text prompts, reference images, or layout maps, which is what enables prompt-driven image generation. Understanding the forward and reverse processes clarifies why these models are slower than GANs but tend to produce more diverse, higher-fidelity, and more controllable results.