Read intro to Diffusion Models

talafek96 / Diffusion-AD-Project

In this project we will attempt to develop a POC for detecting anomalies in images based on the ability or inability of a DDM to reconstruct them.

1 stars 1 forks source link

Read intro to Diffusion Models #2

Closed talafek96 closed 1 year ago

talafek96 commented 1 year ago

https://huggingface.co/blog/annotated-diffusion

Related to epic #1

talafek96 commented 1 year ago

Regarding the algorithm shown, why are we sampling a random t instead of training over every t in a sequence? (since the process of adding noise happens one step after another)

@amitir22

amitir22 commented 1 year ago

in the bullet "Position Embedding":

why is there a math.log(10000)?
why we take the half-dim?
in general, is there an intuitive explanation for what is going on in the method forward(self, time)?
Also, usually we use positional embeddings in recurrent neural networks where the position of the time series matters. What is the purpose of using it in this context?

EliSchwartz commented 1 year ago

Re sampling of a random t, I'm not 100% sure but I believe it's related to stability. You don't want to sequentially perform too many iterations on a single image (it will overfit the particular image).

Re position embedding:

Just a constant for the exp below
Because we have both Sines and Cosines
This encoding of time lets the model know what level of noise to expect

amitir22 commented 1 year ago

what is going on in those lines? EDIT: nvm it's just a cool function.