Here's something a lot of people still don't know: The latest DeOldify doesn't use GANs anymore. And I'm not being cute with terminology- NoGAN isn't used either. We needed something more production worthy and controllable and it just wasn't cutting it. 1/
-
Show this thread
-
Replying to @citnaj
Keyword here being "controllable", I suppose - this is valid in other DL fields too: generative approaches (in CV, in NLP, etc) are very cool on paper / for academia, but frankly unmanageable when verticalized into industries.
1 reply 1 retweet 23 likes -
Replying to @kairosdojo
I can only imagine! To elaborate on "controllable"- We -just- now got to the point where quality is consistent across rendering with a wide range of resolutions. One of many aspects that matter for production.. That aspect is so hard in practice to engineer for.
1 reply 1 retweet 24 likes -
Replying to @citnaj @kairosdojo
Care to elaborate about the difficulties of wide resolution range? More of an engineering or fundamental problem?
3 replies 0 retweets 0 likes
Sure. A combination of training regime itself (how high of a res you can go on with a reasonable batch size) and the architecture itself (e.g spectral norm) can make/break whether or not inference generalizes across resolutions well.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.