t timesteps. 'Noise' represents random perturbations of the pixel values.
The neural network is trained by observing the noise that is added at each timestep and then learning how to go back to the previous timestep's less noisy version:
The input is in the neural net is the noisy image at timestep t + 1, and the labelled output is the less noisy image at timestep t. The network learns to clean up a noisy image into a clearer one.
What it learns about reversing noise is what enables it to generate new images from pure noise during the reverse process.
Training Phase: The model learns by observing how noise is gradually added to real images over many steps.
Generation Phase: Start with random noise and apply the learned denoising process in reverse.
Neural Network: Predicts what noise to remove at each timestep based on the current noisy image.
Markov Chain: Each step only depends on the previous step, making the process mathematically tractable. This memoryless property means the noise addition at timestep t only needs the image from timestep t - 1, not the entire history.
Try your hand at manual denoising! Apply different denoising algorithms to progressively reveal the hidden cat image. The order matters for optimal results!