Course contentsShow
← Machine Learning and Deep Learning
Lesson 1525 of 3,53834. Diffusion ModelsFree lesson

The Markov Chain of Noise Addition

Understand how forward diffusion forms a Markov chain where each step depends only on the previous state.

The Markov Chain of Noise Addition

What you'll learn: How forward diffusion in diffusion models works as a Markov chain where each noising step depends only on the immediately previous state.

Understanding the Markov Property

In diffusion models, we progressively add noise to an image over many timesteps (t=0, t=1, t=2, ... t=T). The Markov property means that to compute the image at timestep t, you only need the image from timestep t-1 — not the entire history of how we got there.

Think of it like a game of "telephone" where each person only hears from the person directly before them. Person #5 doesn't need to know what Person #1 originally said; they only need what Person #4 tells them.

Why This Matters

Mathematically, at each step t, we add a small amount of Gaussian noise to the previous noisy image:

x_t = √(1 - β_t) · x_(t-1) + √(β_t) · ε

where ε is random noise and β_t controls how much noise we add at step t.

The crucial insight: x_t depends only on x_(t-1), not on x_0, x_1, ... x_(t-2). This is the Markov property in action.

This chain structure has two huge benefits:

  1. Mathematical tractability: We can derive closed-form solutions for jumping directly from x_0 to x_t without computing all intermediate steps
  2. Training efficiency: The model can learn to reverse any single step without tracking full histories

By the end of the chain (t=T), the image becomes pure Gaussian noise, having "forgotten" its original structure through this memoryless process.

Key Takeaway: Forward diffusion forms a Markov chain where each noising step depends only on the previous state, not the full history — enabling both mathematical elegance and efficient training of the reverse denoising process.