Live 67: Generative Image-to-Image Translation Networks

In yesterday's live stream—Live 67—I went over various generative image-to-image translation networks, trying to see their main characteristics and learn from them; how they work, what losses they use, which work their built upon. I used this list of networks, which includes, among others, Pix2Pix, CycleGAN, StyleGAN, CUT, and Informative Drawings.

I also showed OpenAI's CLIP, DALL·E, and Microscope, and NVIDIA's Imaginaire.

Browse through the timestamps below to jump into specific parts of the video.

You can spread the word by liking and sharing this tweet.

If this is something that interests you, please let me know on Twitter or, even better, on the Discord community.

Thanks for watching.

See you next week!

Chapters

• 02:02 Introduction
• 11:04 Q&A: Can ML convert sketches into single-line drawings?
• 17:26 Generative image-to-image translation networks
• 23:16 Generative adversarial network architecture
• 25:46 GANs over time
• 30:49 Pix2Pix - Image-to-image translation
• 34:30 CycleGAN - Cycle-consistent unpaired image-to-image translation
• 37:48 CUT - Contrastive unpaired image-to-image translation
• 42:33 DualGAN
• 43:50 CycleGAN
• 46:24 Multi-model unsupervised image-to-image translation
• 59:51 Informative Drawings - Learning to generate line drawings that convey geometry and semantics
• 01:02:40 NVIDIA's Imaginaire
• 01:04:22 CLIP and DALL-E - OpenAI
• 01:06:07 Microscope - OpenAI
• 01:09:01 Outro