Reverse Process as a Learned Denoiser
You'll learn: How diffusion models use a neural network to progressively clean up noise and recover the original data.
The Core Idea
Remember how the forward process gradually added noise until your image became pure Gaussian noise? The reverse process does the opposite: it starts from that noisy mess and learns to remove the noise step-by-step, eventually reconstructing a clean image.
Here's the key insight: at each step of the reverse chain, we need to predict what "slightly less noisy" looks like. A neural network becomes our learned denoiser—it observes the current noisy state and the timestep, then predicts how to remove just the right amount of noise for that step.
Why "Learned"?
Unlike simple filters or traditional denoising, we can't write down an exact formula for reversing diffusion. The relationships are too complex. Instead, we train a neural network (often a U-Net) to learn:
- Input: A noisy image at timestep t
- Output: Information to denoise it (either the noise itself, the clean image, or the mean of the previous step)
- Conditioning: The timestep t tells the network "how noisy" the input is
The Gradual Refinement
Think of it like a sculptor: you don't go from a block of marble to a finished statue in one chisel strike. Similarly, the network doesn't try to jump from pure noise to a perfect image instantly. It takes small, learned denoising steps backward through the Markov chain, gradually revealing structure and detail.
Key Takeaway: The reverse process uses a neural network trained to denoise at every timestep, gradually transforming noise back into realistic data through learned incremental refinements.