Conversation

This paper from

et al at Nvidia combines autoencoders and GANs to bidirectionally transform footage from day to night, winter to summer, real to video-game-ish. Wild! arxiv.org/abs/1703.00848

GIF

Frank A. Krueger

@praeclarum

Nov 8, 2017

I’ve been following pix2pix for awhile and excited to see if they added time to the net (or are just doing frame by frame)

Replying to

and

The architecture described in the paper suggests independent processing of each frame. Different arch from pix2pix; closer to CycleGAN.

5:06 PM · Nov 8, 2017Twitter for Mac

Replying to

and

Ack, and CycleGAN cites this work, so better to say the other way around. :)

Replying to

and

Ah darn. I've been able to get pix2pix into CoreML but CycleGAN has these weird batch operations that don't translate well. Anyway, thanks for the link, can't wait to read!