Skip to content

diffusion

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

diffusion

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
generate_colored_mnist_diffusion.py
generate_colored_mnist_diffusion.py
 
 

README.md

Example usage of diffusion process

python generate_colored_mnist_diffusion.py 
$ python examples/diffusion/generate_colored_mnist_diffusion.py
Epoch 50/50: 100%|██████████| 469/469 [loss=0.0041 avg_loss=0.0049]
Generating 10 images with context tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])..

Generate colored digits

The forward (diffusion) process is gradually adding Gaussian noise to the image at each timestamp:

q(x_{0:T}) = \underbrace{q(x_0)}_{\text{Image}} \prod \underbrace {q(x_t|x_{t-1})}_{\mathcal N_t}

The reverse process is learning the Gaussian transitions to restore the image:

p(x_{0:T}) = \underbrace{ p(x_T) }_{\mathcal N(\mathbf 0,\,\mathbf I)} \prod \underbrace{p_{\theta} (x_{t-1}|x_{t})}_{\mathcal N_t(\mu_\theta,\,\sigma^2\mathbf I)}