I'll let you guys in on a secret: I just figured out something this week that basically makes image to image models do way more for little extra computational or memory cost. Definitely paper-worthy but I'm too busy trying to get this in your hands :)https://twitter.com/thehardmenpath/status/1161670085498552320 …
-
Curious to see what you mean by way more? I guess we'll see on your feed! Is it specific to colorization or could it apply to superresolution/segmentation/etc?
Clarifying- way more performance as in quality of output. Also, sorry about not getting back to you on the attention stuff. Feedback loops are super slow on this stuff and I’ve been repeatedly sidetracked (in a good way)
