The Diffusion Model Forward Process
Diffusion models learn to generate data by first defining how to destroy it. The forward process slowly adds noise to a sample over many steps until nothing remains but pure noise.
Step by step corruption
- Start with a real example, such as a clean image.
- At each step add a small amount of Gaussian noise according to a fixed schedule.
- After enough steps the sample is indistinguishable from random noise.
Key properties
- The forward process has no learnable parameters. It is a fixed recipe.
- Thanks to the math of Gaussians, you can jump to any noise level in one step rather than iterating, which makes training efficient.
- The noise schedule controls how fast corruption happens and strongly affects final quality.
Why define destruction first
By knowing exactly how data was corrupted, the model has a clear target to reverse. The forward process supplies training pairs of a noisy sample and the noise that was added, which the reverse process learns to predict.
Key idea
The diffusion forward process is a fixed parameter free schedule that adds Gaussian noise to data step by step until it becomes pure noise, creating the training targets for the reverse model.