Hello! I've made a framework of a diffusion model which diffuses a one-hot encoded representation of a letter, and tries generating a 28x28 image of that letter. However, I have noticed that the network does not seem to be learning how to decode the diffusion, and I have no idea if I did something wrong in the diffusion process (the function name in the project is add_noise_to_onehot). Can someone please take a look at it and debug it (the training instance has everything). Also all of the other sprites in the project are not used except the Net3 sprite (which has everything).
Thanks!
Also it just diffuses text and feeds it into a otherwise normal neural network, if that helps.
Link to project:

https://turbowarp.org/880999122