GANs, Diffusion Models, Generative Tasks (txt2img, img2img, inpainting)

[sarthak247]

Greetings everyone,
Inspired by #19 , me and my fellow collaborators have also outlined a course curriculum for our section but we would like to have some inputs and feedback from the HF team before we finalise it and start working on it. This is our chosen structure so far.

INTRODUCTION

• What are generative vision models and how do they differ from other models?
• Different types of generative models/tasks?
• Prerequisites and resources to help

GANS & VAEs

• VAEs theory (Theory)
• Idea behind GANs, generator and discriminator (Theory & code)
  • DCGAN as the main implementation
• Simple explanation, showcase and external resources:
  • StyleGAN
  • CycleGAN
  • VQGAN

Diffusion models

• Theory of diffusion models and how they differ from GANs (limitations of GANs)
• Evolution/ what made it work diffusion models DDPM, latent diffusion
• Using stable diffusion
  • Basic structure of SD
  • How to use txt2img, img2img, inpainting
• Simple explanation, showcase and external resources
  • Dreambooth
  • LoRA, show how to use, link to fine-tuning yourself
  • ControlNet, show how to use

PRACTICAL APPLICATIONS & CHALLENGES

• Real-time Constraints and Privacy Concerns
• Bias concerns

---

CC: @hwaseem04 , @mattmdjaga, @charchit7
BCC: @johko , @lunarflu , @merveenoyan

We would like to know if this course structure is suitable or do there need to be changes for this. In particular we are interested to know:

• Do we need to create separate files for each topic or squish it together into a big jupyter notebook? (essentially structural help)
• How much code is too much? Like, for GANs, we've thought to stick for DCGAN for code from scratch and for the rest just refer to the pre-trained models and showcase how to use them directly.
• Is the chosen curriculum sufficient or do we need to add more to it? Also, if it's too much, then what can be truncated or removed from it considering that we also have an existing diffusers course out there.

Thanks,
Sarthak (and @hwaseem04 , @mattmdjaga, @charchit7)

[lunarflu]

Great job! 🤗 A few comments:

• I like the emphasis on theory always being first, always ensuring we know WHY a certain thing is important to learn. And then building on past ideas and emphasizing what has evolved - very nice!
• As for "amount of code", I don't think we have a hard ceiling. Use what you think is helpful, exercise common sense, and it should be fine 😌💪

[jere357]

Hello, i would like to help on this section, my discord username is cropinky. My introduction to adversarial learning was ESRGAN and i think it would be a cool part of this section.

[johko]

hey @sarthak247

thanks for coming up with this curriculum. I think it covers everything we need for this course. And I don't have much to add.

With our new folder structure where .mdx and notebook files live in separate repos you also have a bit more flexibility in dividing the coding heavy parts from the theoretical ones. Definitely feel free to not squish it all into one file ;)

[charchit7]

Hey @merveenoyan, @sayakpaul would love your inputs on this one :)

[arkajyotimitra]

Great outlook to cover about GANs and diffusion models.
Since diffusion model has become such a vast concept in itself and there is already a diffusers course to explore both the length and breadth of it. The simplistic introduction towards the concept is apt.
One thing that might be helpful to look at is the association of diffusion models with physics (the place from where it originated). Some difference between score-based and energy-based diffusion models. To that degree these resources might may come handy:

• a video lecture by Jascha Sohl-Dickstein
• a video explanation of the diffusion models evolution by Yang Song
• an insightful blog on different perspectives and their associations around diffusion models by Sander Dieleman

This might get too deep so I will leave it at your discretion to use it or just have fun reading/listening these sources 🤗. I just wanted to share as I enjoyed them and these gave me more insights about diffusion models as a whole.