Topics to Look For¶
- Generative Adversarial Network (GAN)
- Diffusion Models
Resources¶
- Generative Adversarial Networks (GAN) (from Dive into Deep Learning).
- Introduction to Diffusion Models for Machine Learning provides an overview of how diffusion models work (without going too deep into the rather hairy math) and presents several of the current (2024) state-of-the-art models for image and video generation that use diffusion.
Examples (GAN)¶
- This Person Does Not Exist — An early popular example of GANs from 2019. It displays a randomly generated photorealistic picture of a person (who does not really exist). Reload the page to get a new image.
- This X Does Not Exist — Several sites similar to This Person Does Not Exist popped up, and this site catalogs them. Many are now gone from the internet, with broken links, but you can see the range of examples people created there. A few examples still online:
- This Automobile Does Not Exist
- This Chair Does Not Exist — Applies a GAN to the task of generating a 3D model, rather than a 2D image.
- GauGAN — An experimental tool that generates photorealistic images based on a rough “sketch” drawn and updated in realtime.
Note that most applications of GANs have already (within 5 or so years) been superceded by newer architectures, including diffusion models and transformers.
Supplemental¶
- GANs in Computer Vision from AI Summer introduces the idea of the GAN and presents several variants that have been developed in advances over time.
- How Diffusion Models Work: The Math from Scratch from AI Summer does a nice job presenting the details of diffusion models.
- Chapter 15: Generative Adversarial Networks in Understanding Deep Learning (very detailed and math/theory-heavy).
- Chapter 18: Diffusion Models in Understanding Deep Learning (very detailed and math/theory-heavy).