A video about using AI to generate youtube thumbnails. I explore the classic GAN method and compare it with a newer method called diffusion. One turns out to be better than the other!
Reviewed by Andrew Carr:   / andrew_n_carr  
Disclaimer: All thumbnails were deleted after use. I do not aggregate youtube data.

The losses are not exactly inverse for the generator and discriminator because the two are not trained on the same data.

LINKS
Twitter:   / max_romana  
Discord:   / discord  
Patreon:   / emergentgarden  
The life engine: https://thelifeengine.net


SOURCES
Original GAN Paper: https://proceedings.neurips.cc/paper/...
Face interpolation:    • StyleGAN2 Interpolation Loop  
BigGAN Paper: https://arxiv.org/abs/1809.11096
https://thispersondoesnotexist.com
Flower Gan:   / 1527890938386857984  
Katydid:    • The Katydid (Leaf Bug)  
Mantis:    • Praying Mantis Hunts a Cricket  

Diffusion Paper: https://arxiv.org/abs/2006.11239
Diffusion beats GANs: https://arxiv.org/abs/2105.05233?curi...
Blog Post: https://gretel.ai/blog/diffusion-mode...
Diffusion explanation:    • Diffusion Models | Paper Explanation ...  
Diffusion Visualization:   / 1537042940475883520  
Water Diffusion:    • demo - (hot and cold water with food ...  

Dall-E 2: https://openai.com/dall-e-2/
Imagen: https://imagen.research.google/
CogView: https://arxiv.org/abs/2105.13290
Parti: https://parti.research.google/


TIMESTAMPS
(0:00) Intro
(0:32) The Goal
(1:11) The Data
(2:15) Latent Image Generators
(3:03) GANs
(4:35) GAN training
(7:45) Diffusion
(8:55) Diffusion training
(10:43) ☆Generated thumbnails☆
(13:58) Diffusion beats GANs
(15:27) Conclusion
(16:28) Outro


MUSIC
   • Closed Circuits