Forward Diffusion Process in DDPM
What you'll learn: How DDPM systematically destroys image structure by adding Gaussian noise in T small steps using a predefined schedule.
The Core Idea
The forward diffusion process is like watching a photograph fade into static television noise—but in a mathematically controlled way. Over T timesteps (often 1000), DDPM gradually adds small amounts of Gaussian noise to your data until it becomes pure noise indistinguishable from a standard normal distribution.
Think of it like this: imagine dropping a drop of ink into water. At first, you see the ink clearly. After stirring once, it's slightly dispersed. Keep stirring 1000 times, and eventually the ink is completely diffused—you can't tell where it started.
How It Works
At each timestep t from 1 to T, the forward process takes the slightly-noisy image from step t-1 and adds a bit more Gaussian noise. The amount of noise added is controlled by a fixed variance schedule (β₁, β₂, ..., βₜ), predetermined before training.
Key properties:
- Each step is a Markov process: timestep
tonly depends ont-1, not the entire history - The noise variance typically follows a schedule that starts small and grows (linear, cosine, etc.)
- By timestep T, the original data structure is completely destroyed—you're left with pure Gaussian noise
The forward process requires no learning—it's entirely deterministic and defined by your variance schedule choice.
Key Takeaway: The forward diffusion process systematically corrupts data into noise over T steps using a fixed variance schedule, creating a tractable path from data to noise that the reverse process will learn to undo.