Sample Creation

Overview

The sampling process in diffusion models involves manipulating noise and iteratively modifying the initial noise to generate new data samples that follow the distribution learned during the model training phase. This process allows us to visualize the generation of images or data points and observe how the model creates new samples based on the learned patterns from the training data.

Sampling

In this lesson, we’ll discuss sample creation, its details, and how it works across multiple different iterations of neural networks. Once the neural network is trained, sampling comes into play.

In the figure above, we have a noise sample. We passed it to our trained neural network (NN). The NN understood the characteristics of a character and predicted noise in the image. The last step is to subtract the predicted noise from the noise sample to see our image more clearly. Realistically, that's just a noise prediction, and it doesn’t fully remove all the noise, so ...