Course contentsShow
← Machine Learning and Deep Learning
Lesson 1543 of 3,53834. Diffusion ModelsFree lesson

Reverse Process: Learning to Denoise

Training a neural network to predict and remove noise, reversing the diffusion process step by step.

Reverse Process: Learning to Denoise

What you'll learn: How diffusion models train a neural network to progressively remove noise and recover clean data.

The Core Idea

You've seen how the forward process gradually adds noise to data until it becomes pure Gaussian noise. Now comes the magic: can we learn to reverse this process? The reverse process trains a neural network to act as a smart denoiser that predicts and removes the noise added at each diffusion step.

Think of it like watching a video of ink dispersing in water, then playing it backwards. The neural network learns to "rewind" the diffusion, taking a noisy image and predicting what slightly-less-noisy version came before it.

How It Works

At each timestep t, the network receives:

  • A noisy image (current state)
  • The timestep t itself (as conditioning information)

The network outputs a prediction of either:

  • The noise that was added, or
  • The less-noisy previous state, or
  • The original clean image

During training, we have access to both the noisy version and what was added, so we can compute a loss (typically MSE) between the network's prediction and the true noise/image. The network learns patterns of what noise "looks like" at different timesteps.

Why Time Conditioning Matters

The same noisy image could have come from different points in the diffusion chain. Telling the network the timestep t helps it understand how much noise to remove—a lot near the end of diffusion, less at earlier steps.

Key Takeaway: The reverse process trains a neural network to predict and remove noise conditioned on timestep, effectively learning to undo the forward diffusion process one step at a time.