The noise-to-image training process
We have the solution to add noise to the image, which is known as forward diffusion, as shown in Figure 4.6. To recover an image from the noise, or reverse diffusion, as shown in Figure 4.6, we need to find a way to implement the reverse step p θ(x t−1| x t). However, this step is intractable or uncomputable without additional help.
Consider that we have the ending Gaussian noise data, and all those noise step data in hand. What if we can train a neural network that can reverse the process? We can use the neural network to provide the mean and variance of a noise image and then remove the generated noise from the previous image data. By doing this, we should be able to use this step to represent p θ(x t−1| x t), and thus recover an image.
Figure 4.6: Forward diffusion and reverse process
You may ask how we should calculate the loss and update the weights. The...
