azad-academy / denoising-diffusion-model

A simple guide to diffusion models. Helpful in understanding the concept and practicing with the method.

Denoising Diffusion Model

This is a simple guide to diffusion models.

Requirements

• Python >= 3.7
• pytorch >= 1.6
• CUDA Toolkit
• GPU

The tutorial 📃

On Medium:

https://azad-wolf.medium.com/diffusion-models-made-easy-8414298ce4da

On Substack:

https://azadwolf.substack.com/p/diffusion-models-made-easy

References

Sohl-Dickstein, J., Weiss, E. A., Maheswaranathan, N., & Ganguli, S. (2015). Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585.

Max Welling & Yee Whye Teh. “Bayesian learning via stochastic gradient langevin dynamics.” ICML 2011.

Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. arXiv preprint arXiv:2006.11239.

Prafulla Dhariwal, Alex Nichol, Diffusion Models Beat GANs on Image Synthesis, arXiv: 2105.05233

denoising-diffusion-model's Issues

model not converge

Hi, thanks for your excellent project !

I am new to diffusion models and I am now trying it on my own dataset (100 images of a specific senario). After being trained for 60k iterations with batch size 8, the model produces real images but with much noises. I have no idea how to debug. How can I improve the training of the model. Some tricks to share ?

Appreciate any help !

Embedding question in model

The output in the model is multiplied by an embedding (for each time as far as I understand). Is this a perturbation added for each time - perhaps r in the original diffusion paper? Trying to figure out how this fits in.

Thanks!

how can I generate image with classes condition information?

thanks your code , it is most clearly I see, and help me understand what is the ddpm

and I want to know how to condition generate with ddpm , would you update your blog ?

Typo in equation in colab notebook

I believe the second q should be x_{0} instead of x_{t-1} in the colab notebook

$$
q(\mathbf{x}{t}\mid\mathbf{x}{t-1}) = \mathcal{N}(\mathbf{x}{t} ; \sqrt{1-\beta{t}}\mathbf{x}{t-1},\beta{t}\mathbf{I})
$$

Substituting $\alpha_{t}=1-\beta_{t}$ and $\bar{\alpha}{t} = \prod{s=1}^{t} \alpha_{s}$

$$
q(\mathbf{x}{t}\mid\mathbf{x}{t-1}) = \mathcal{N}(\mathbf{x}{t} ; \sqrt{\bar{\alpha}{t}}\mathbf{x}{t-1},(1-\bar{\alpha}{t})\mathbf{I})
$$