First Notes about Simple Diffusion Model

The notes is based on the online course of Hugging Face. 

 ### Basic Philosophy of a Simple Diffusion Model
Training a simple diffusion model:
- A simple diffusion model can be divided into 2 types of processes, to name it  forward and backward propagation. 
- Training data: inputs of pictures 
- Forward Propagation: adding noises into the pictures, making the pictures tends to be complex and blurred in vision, until reaching Gaussian noise (pdf in normal distribution) 

- Backward Propagation: denoise the blurred pictures back to the original pictures, and updating weights during the processes. 

After finishing training the models:
- The inputted test data will be used to generate new pictures by the model.

- The forward propagation process is named Corruption.

### Sampling
- sampling is the method used to generate images from random noised pictures. 
- After the first trial, feed a small proportion of the structure of the prediction into the input set, and restart the process of prediction 
- After a few trial, a clear prediction will be made since we have a better starting point each trial 

The above sampling is based on a simplified U-Net 

The above sampling making better prediction is based on U-Net2D

赛博猫咪图：