habib61 / denoising_diffusion_models

PyTorch Implementation of Denoising Diffusion Probabilistic Models

PyTorch Implementation of Denoising Diffusion Probabilistic Models

Papers implemented

• Denoising Diffusion Probabilistic Models
• CLASSIFIER-FREE DIFFUSION GUIDANCE

Models

https://huggingface.co/singhdan/diffusion_models

Results

---

Unconditional Diffusion Model

Sample images

Here's a gif I made from the last 20 training epochs :)

---

Conditional Diffusion model

Sample images

All of them are frogs 🐸(it's one of the labels in the CIFAR10 dataset), because why not :), although maybe a horse could have been a nice choice too.

This time we have done things 4 ways, since we have CFG(Classifier free guidance) and EMA(Exponential Moving Average)

🚫EMA 🚫CFG

🚫EMA ✅CFG

✅EMA 🚫CFG

✅EMA ✅CFG

Note: In the CFG paper the authors mentioned heavy saturation when CFG levels were too high(i.e = 3), we are able to reproduce that same effect in the above picture.

Clearly the results with EMA are much better generally, however CFG helps with more stable looking images :)

---

Dataset used

• Landscapes dataset (for unconditional model)
• CIFAR-10 Resized (for conditional model)

Helpful resources

• https://github.com/lucidrains/denoising-diffusion-pytorch
• https://github.com/hojonathanho/diffusion
• https://www.youtube.com/watch?v=TBCRlnwJtZU
• https://theaisummer.com/unet-architectures/