This is a baseline 'reproduction' of the work Generative Modeling by Estimating Gradients of theData Distribution by Song et al.
Here I use the MNIST dataset with a modified UNET network. The UNET decoder uses Conditional Instance Normalization ++
from the paper to decode based on the input noise. This is not added to the encoder due to technical reasons.
No noise added | Noise added |
---|---|
MNIST | CIFAR10 |
---|---|
@misc{qbeer,
author = {Alex Olar},
title = {Noise Conditional Score Network},
howpublished = {GitHub repository},
month = {February},
year = {2023},
url = {https://github.com/qbeer/noisy_diffusion}
}